Critical · no confirmed exploitation TLP:CLEAR Compiled 20 Nov 2026 · disclosed 17 Nov · Oracle and OpenJDK patched 17 Nov 21:00 UTC · not in CISA KEV (catalogue v2026.11.18, 1,786 entries) · NVD record published 17 Nov 22:18 UTC

Java4Shell: CVE-2026-114938

Intelligence brief for SOC and vulnerability-management teams. A heap out-of-bounds write in OpenJDK’s AVX-512 Base64 decoder, disclosed on 17 November 2026 and patched the same day — demonstrated, by its discoverers, as unauthenticated remote code execution against default installs of Jira, Confluence, Jenkins, Keycloak, Artifactory and PingFederate. No exploitation in the wild has been reported. This page tracks the vulnerability, the patch rollout, and what independent labs have and have not reproduced.

Bottom line

  1. The bug is real, and the trigger is trivial. Sixty-four A characters followed by a single = make the HotSpot JVM’s hardware-accelerated Base64 decoder write 62 bytes past the end of a 50-byte heap array. The write happens before Java rejects the input as invalid. The discoverers turned it into a heap leak (“JavaBleed”) and then into remote code execution.
  2. RCE is demonstrated by the discoverers, not yet independently reproduced. Calif claims pre-authentication RCE against six widely deployed products in their default configurations; Wiz says it proved remote exploitation feasible against Jenkins, Confluence and Apache Tomcat. Other labs have reproduced the overflow, JVM crashes and a local heap disclosure — but not remote code execution. Oracle rates the flaw 8.1 (CVSS 3.1, high attack complexity); Calif rates it 9.2 (CVSS 4.0).
  3. Patching is under way but uneven. Oracle’s fixed builds shipped on 17 November, as did Amazon Corretto and Azul Zulu. Red Hat has shipped fixed builds for its Build of OpenJDK (portable and Windows), and all six RHEL errata shipped on 19 November; the Red Hat container images and the Red Hat Build of Keycloak remain affected. Eclipse Temurin has now published fixed binaries for all four lines on x64 Linux, Microsoft has shipped its fixed builds, and BellSoft has confirmed fixed CSPU builds (17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1) that include the patch. Okta has published fixed versions for its Access Gateway and agents (19 November). Two JVM flags close the exposure where patching must wait — but one of them, -XX:UseAVX=2, stops OpenJDK from starting on ARM, so use -XX:-UseBASE64Intrinsics on mixed estates.
  4. The hard part is inventory, not the patch. The flaw is in the Java runtime, not in any application or library, so dependency scans will not find it — and runtimes bundled inside third-party products, container images and appliances may not appear in software inventories at all. Three conditions must all hold: an OpenJDK build from 17.0.6 onward, an x86-64 processor with AVX-512 VBMI, and an application that decodes attacker-controlled Base64.

Status board

Where this stands, 20 November

Vendor advisory
Shipped
OpenJDK advisory and Oracle November Critical Patch Update, 17 Nov 21:00 UTC (13:00 PT) · 11 Java SE fixes in the update, all remotely exploitable without authentication
Fixed versions
Oracle .1 builds
17.0.21.1 · 21.0.13.1 · 25.0.5.1 · 27.0.1.1, plus 8u513-perf for the Java 8 Enterprise Performance Pack
CVE record
Published 17 Nov
Published 17 Nov 22:18 UTC with Oracle’s CVSS 3.1 score of 8.1; status “Undergoing Analysis”; NVD has not added its own score yet
In CISA KEV
No
Catalogue v2026.11.18, 1,786 entries · CISA issued an alert on 17 Nov urging immediate patching
Exploited in the wild
None confirmed
GreyNoise has recorded 197 probing sessions from seven IPs, no successful intrusion (tracker read 08:58 UTC, 20 Nov) · Tenable, Sophos, Qualys and Cisco Talos all report no exploitation as of 17 Nov
Public exploit
Trigger only
Calif published a safe reproducer that prints REPRODUCED; its product-specific chains are withheld · 55 GitHub PoCs counted by cve.now as of 18 Nov, first 18 Nov 02:15 UTC — two verified and linked below (arkrwn, crash and heap-leak reproducer; ikow, trigger, chain, scanners and detection rules)
Distributions pending
Temurin Win; macOS 21/25; Red Hat containers
Temurin has published fixed binaries for all four lines on x64 Linux (17.0.21.1+2, 21.0.13.1+2, 25.0.5.1+2, 27.0.1.1+4) and the fixed Java 17 build for macOS x64 (19 Nov); Windows builds and the Java 21 and 25 macOS builds are still in testing. Microsoft shipped its fixed builds on 18 Nov; BellSoft confirmed fixed CSPU builds 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1. Red Hat has shipped fixed builds for its Build of OpenJDK (17.0.21.1, 21.0.13.1, 25.0.5.1, 18 Nov), and all six RHEL errata shipped on 19 Nov (java-17 for RHEL 8 and 9; java-21 for RHEL 8, 9 and 10; java-25 for RHEL 9 and 10). The Red Hat container images and the Red Hat Build of Keycloak remain affected; Okta published fixed agent versions on 19 Nov
 
Detection is behavioural
No network IOCs have been published — hunt JVM crashes, unexpected restarts and the 65-byte trigger in request logs, not IP addresses
Scores, by scorer
9.2 / 8.1 / 9.8
Calif 9.2 (CVSS 4.0) · Oracle and OpenJDK 8.1 (CVSS 3.1, high attack complexity) · Red Hat 9.8, now with its vector published (attack complexity low) — do not merge; no CVSS 4.0 vector has been published

Confidence labels used below: High vendor or government primary, or independently reproduced · Medium single-researcher claim corroborated in part · Low uncorroborated claim, or conflicting reports.

↑ Back to contents

At a glance

Java4Shell in brief

A memory-corruption flaw in the Java runtime itself, disclosed and patched on 17 November 2026. 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 NovemberOkta publishes fixed versions; Red Hat completes the RHEL errata; Temurin ships macOS Java 17Show updateHide update

The patch rollout reaches the identity layer, and the last RHEL gaps close. Still no exploitation in the wild, and the flaw is still not in CISA KEV.

Okta publishes fixed versions. Okta’s security notice, updated 19 November at 13:51 PST (21:51 UTC), says the necessary patches are actively rolling out across Okta’s production environments; Auth0 environments are not affected. Customers should upgrade: Okta Access Gateway to 2026.45.1 or greater, LDAP Agent to 5.26.3 or greater, On-Premises Provisioning Agent to 3.3.4 or greater, and RADIUS Agent to 2.30.0 or greater.

Correction: Red Hat’s RHEL errata are complete. All six RHEL errata shipped on 19 November: java-17-openjdk 17.0.21.1.2 for RHEL 8 (RHSA-2026:100563) and RHEL 9 (RHSA-2026:100564); java-21-openjdk 21.0.13.1.2 for RHEL 8 and 9 (RHSA-2026:100567) and RHEL 10 (RHSA-2026:100568); and java-25-openjdk 25.0.5.1.2 for RHEL 9 (RHSA-2026:100571) and RHEL 10 (RHSA-2026:100572). This corrects the “coverage is still partial” line in the 19 November update and the vendor table below; the Red Hat container images and the Red Hat Build of Keycloak remain affected.

Correction: Temurin has shipped its fixed Java 17 build for macOS. The fixed JDK 17.0.21.1+2 binaries for macOS x64 (pkg and tar.gz) were published on 19 November at 23:02–23:03 UTC. This corrects the “Windows and macOS builds are still in testing” line in the 19 November update and the status board and vendor table below: Windows builds and the Java 21 and 25 macOS builds are still in testing.

Still no exploitation, still not in KEV. GreyNoise’s count now stands at 197 sessions from seven IP addresses, with no successful intrusion seen on its sensors; its tracker was read at 08:58 UTC on 20 November. A seventh IP, 185.184.192.254 (WorldStream, Netherlands), first tracked at 20:48 UTC on 19 November, is probing Jenkins endpoints (/api/json, /crumbIssuer/api/json, /whoAmI/api/json, /cli). CISA’s KEV catalogue is still v2026.11.18 (1,786 entries) and CVE-2026-114938 is still not in it. No confirmed exploitation and no named victim.

Sources: Okta security notice; Red Hat RHSA advisories; GreyNoise tracker; CISA KEV; Adoptium Temurin 17 release.

↑ Back to contents

Update · 19 NovemberTemurin finishes x64 Linux, Microsoft ships, and the first product advisories landShow updateHide update

The patch rollout moved a long way overnight: the two distributions that were still pending on 18 November have now shipped, and the first vendors whose products Calif named have published their own guidance.

Patch rollout. Eclipse Temurin has published fixed binaries for all four lines on x64 Linux — 17.0.21.1+2, 21.0.13.1+2, 25.0.5.1+2 and 27.0.1.1+4 (18 November, evening UTC); Windows and macOS builds are still in testing. Microsoft shipped its fixed builds on 18 November — 17.0.21.1, 21.0.13.1 and 25.0.5.1 — with CVE-2026-114938 named in the release notes. BellSoft published a blog on 18 November (updated 19 November) confirming Liberica CSPU builds 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1 include the patch for CVE-2026-114938; its fixed-build tags are 17.0.21.1+6, 21.0.13.1+6, 25.0.5.1+5 and 27.0.1.1+5. Red Hat has shipped fixed builds for its Build of OpenJDK — 17.0.21.1, 21.0.13.1 and 25.0.5.1, released 18 November (RHSA-2026:100565, RHSA-2026:100566, RHSA-2026:100569, RHSA-2026:100570 and RHSA-2026:100573, portable and Windows builds). RHEL RPMs have also begun shipping — java-17-openjdk 17.0.21.1.2 for RHEL 8 (RHSA-2026:100563) and java-21-openjdk 21.0.13.1.2 for RHEL 10 (RHSA-2026:100568) — though coverage is still partial, and its bulletin RHSB-2026-018 remains open.

First product advisories. Siemens published advisory SSA-272816 (18 November): Industrial Edge Management Pro V1 and V2 and Industrial Edge Management Virtual are affected in all versions, with no fix available yet; Siemens advises blocking direct internet access to them, or shutting instances down. The Industrial Edge Device Kit and Industrial Edge Management Cloud are not affected. Keycloak published its own guidance (18 November): its 26.8.2-1 and 26.8.2-2 x86-64 container images bundle an affected OpenJDK 21.0.13 build, and the Keycloak Operator image (built with OpenJDK 25) is also exposed; the interim mitigation is -XX:-UseBASE64Intrinsics with a restart. On 19 November Keycloak updated the post: patched images are now available, including 26.8.2-3, built on Red Hat OpenJDK 21.0.13.1. Apache Shiro published its own guidance (18 November): Shiro itself is not vulnerable, but its CookieRememberMeManager Base64-decodes the remember-me cookie through java.util.Base64, so Shiro applications on a vulnerable JDK may be exposed; the fix is the JDK update, with -XX:-UseBASE64Intrinsics as the interim mitigation. A 19 November SecurityWeek round-up adds further vendor impact: Cisco confirms more than a dozen products affected, with no fixes or workarounds yet; Elastic says every Elasticsearch release from 7.17.3 onward, and Logstash from 8.6.2, is affected; Atlassian recommends disabling AVX-512 for its Data Center products until patches arrive; and Jenkins advises updating to the November Java releases.

Correction: what Wiz actually proved. Wiz’s own text says it proved remote exploitation feasible against Jenkins, Confluence and Apache Tomcat — not Jira, Confluence and Jenkins, as this page has stated. Jira sits in Wiz’s broader “proven vulnerable applications” list, which reproduces Calif’s six. The summary sections above are corrected; the dated 17 November entries below are left as written. Apache Tomcat is now named in the demonstrated-applications section.

Correction: Red Hat has shipped fixed builds. Red Hat’s Build of OpenJDK fixed builds — 17.0.21.1, 21.0.13.1 and 25.0.5.1 — were released on 18 November (RHSA-2026:100565, RHSA-2026:100566, RHSA-2026:100569, RHSA-2026:100570 and RHSA-2026:100573). RHEL RPMs have also begun shipping — java-17-openjdk 17.0.21.1.2 for RHEL 8 (RHSA-2026:100563) and java-21-openjdk 21.0.13.1.2 for RHEL 10 (RHSA-2026:100568) — though coverage is still partial: other RHEL and Java combinations, the Red Hat container images and the Red Hat Build of Keycloak remain affected. This corrects the statement in this morning’s update that Red Hat had no fixed build; the summary sections and the vendor table below are corrected.

Correction: BellSoft has confirmed its fixed builds. BellSoft’s blog, published 18 November and updated 19 November, confirms Liberica CSPU builds 17.0.21.1, 21.0.13.1, 25.0.5.1 and 27.0.1.1 include the patch for CVE-2026-114938. This corrects the statement in this update that the fix was only inferred from the version numbers; the summary sections and the vendor table below are corrected.

Scores. Red Hat’s CVE page now publishes the vector behind its 9.8: CVSS:3.0/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H. Attack complexity low, against Oracle’s high, is the whole disagreement between the two scores. Red Hat’s statement also says public research has demonstrated working exploits against multiple widely deployed Java applications in their default configurations. The 9.8 is still described as preliminary in reporting.

Still no exploitation, still not in KEV. No confirmed exploitation in the wild and no named victim. CISA’s KEV catalogue moved to v2026.11.18 (1,786 entries) and CVE-2026-114938 is still not in it. GreyNoise has now recorded 192 sessions from six IP addresses as of 19 November 15:00 UTC — a sixth, 13.212.144.112 on Amazon EC2 in Singapore, first tracked at 13:48 UTC — still with no successful intrusion seen on its sensors; Previdian logged nine attempts from six IPs on 18–19 November, also without confirming any compromise. Cloudflare’s Cloudforce One reported that a source IP first scanned a target with a User-Agent that referenced Java4Shell and returned about 17 hours later with a Java4Shell attempt disguised as Chrome, and separately that it had seen ARTEX and Java4Shell traffic from a single IP address for the first time — neither observation establishes that ARTEX sent the Chrome-disguised request, that either attempt succeeded, or that the two came from a common operator. The count of GitHub PoCs rose from 24 to 55 (cve.now, 18 November); none of the new repositories has been verified, and none is linked here.

Sources: Adoptium API; Microsoft Build of OpenJDK; BellSoft Liberica blog; Siemens SSA-272816; Keycloak; Apache Shiro; Red Hat CVE page; Red Hat RHSA advisories; Wiz; CISA KEV; NVD; cve.now; Cyber Security News; SecurityWeek; GreyNoise tracker.

↑ Back to contents

Update · 18 NovemberVendor fixes land unevenly; GreyNoise sees probing; independent labs reproduce the crash, not the shellShow updateHide update

The first full day after disclosure splits into two stories: the patch rollout, which is fast for some vendors and only now starting for others, and the exploitation picture, where every independent reproduction so far stops at a crash or a local heap leak.

Patch rollout. Amazon Corretto shipped fixed builds on 17 November (17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2, 27.0.1.1.4), with CVE-2026-114938 named in the changelogs. Azul Zulu shipped fixed free builds (17.70.201, 21.54.201, 25.38.201, 27.30.201). SAP released SapMachine .1 builds, though its release pages do not say what they fix. Red Hat published bulletin RHSB-2026-018 but has not yet released fixes for RHEL 8, 9 and 10 or its OpenJDK 17, 21 and 25 builds, saying they are being expedited; the bulletin also names the Azure instance families most likely affected (Dsv5, Dsv6, Dasv6, Dasv7 and related D-series v5–v7 SKUs). Eclipse Temurin has no fixed build yet — its current 17.0.21, 21.0.13 and 25.0.5 builds are affected. The Microsoft Build of OpenJDK and BellSoft Liberica had no fixed builds listed when The Hacker News checked on 18 November.

Probing has begun. GreyNoise is monitoring scanning and exploitation attempts and has detected requests probing for the flaw, with no confirmed successful attack. Cisco Talos said on 17 November that it had not seen in-the-wild exploitation but expected probing to increase after disclosure.

Independent reproductions. Blackpoint Cyber’s Adversary Pursuit Group reproduced JVM crashes under normal runtime configurations, including through repeated unauthenticated HTTP requests against a local Spring Boot application using Spring Security, and reproduced Calif’s writer/reader heap-disclosure technique under controlled local conditions — but did not independently reproduce remote code execution. Harness (project Java2Shell) reproduced the overflow and a JVM crash with small fixed TLABs (SIGSEGV in the G1 collector) and did not pursue code execution; its verdict is “pre-auth heap corruption with a demonstrated denial-of-service”. Upwind reproduced the overflow on an AWS c6i instance (Intel Ice Lake) and confirmed that -XX:-UseBASE64Intrinsics stopped it on every affected version tested. Security firm Immersive said it had built its own working exploit. Equixly calls RCE unproven and a JVM crash the realistic outcome.

Government response widens. South Korea’s Ministry of Science and ICT and the Korea Internet & Security Agency (KISA) urged urgent checks on 18 November, posting an advisory on the BohoNara website and directing major information-and-communications infrastructure facilities, ISMS-certified companies and companies that have reported a CISO to act. New Zealand’s NCSC issued its own alert, expecting most Java applications on affected runtimes to be impacted.

First GitHub PoCs. cve.now counts 24 proofs of concept on GitHub as of 18 November, the first appearing at 02:15 UTC (Ghost-Bx1/java-exploits, r3dvei1/java4shell). Two have since been verified and linked in the PoC section below: arkrwn/PoC-CVE-2026-114938-CrashLeak, a crash and heap-leak reproducer, not an RCE shell; and ikow/java4shell-poc, the most complete public toolkit so far — trigger, staged chain, per-target modules, scanners and detection rules. The rest have not been verified individually.

Evening: Temurin ships its first fixed binary; NVD record published. Eclipse Temurin has named the fixed versions for CVE-2026-114938 — 17.0.21.1+2, 21.0.13.1+2, 25.0.5.1+2 and 27.0.1.1+4 — and, per its release tracker, published its first binary, JDK 27.0.1.1 for x64 Linux; the remaining builds are in testing and will be published as testing completes. Red Hat still has no fixed build. Separately, the NVD record was published on 17 November at 22:18 UTC with Oracle’s CVSS 3.1 score of 8.1 — the “still RESERVED” line in the status board was already out of date when this page was compiled, and is corrected above. NVD has not added its own score yet, and CISA’s SSVC assessment records exploitation as none. Singapore’s CSA issued alert AL-2026-166 urging prompt patching; it adds no new technical facts beyond the Oracle advisory.

Sources: Censys advisory; The Hacker News; ThreatDefence; GreyNoise via Malware News; Blackpoint Cyber; Harness; Upwind; cve.now; Korea MSIT/KISA via FN News; NCSC NZ via CVETodo; Adoptium; CSA Singapore AL-2026-166; NVD.

↑ Back to contents

Update · 17 NovemberDisclosure day: leaked details, same-day patches, a CISA alert, and a 9.2 that Oracle scores 8.1Show updateHide update

The flaw arrived in public almost all at once: a leak, a patch, a technical write-up and a government alert inside a single day.

Before the day. Oracle published a pre-release notice for its November Critical Security Patch Update on 16 November at 08:00 UTC, covering 11 Java SE fixes, all described as remotely exploitable without authentication, with a maximum CVSS 3.1 score of 8.1 — but naming no CVE identifiers. Calif says details of Java4Shell leaked ahead of the coordinated disclosure; it has not said how.

17 November, 21:00 UTC (13:00 PT). Oracle shipped its November Critical Security Patch Update — 655 vulnerabilities across its portfolio, 62 of them rated critical — and the OpenJDK Vulnerability Group published its advisory (last updated 17:44 UTC), naming CVE-2026-114938 in its risk matrix under hotspot/compiler at 8.1. Oracle’s advisory lists the affected builds and credits Quang Luong of Calif and Albert Yang with the report. Oracle’s advisory itself publishes no CVE identifiers for the Java fixes; the OpenJDK advisory is what ties CVE-2026-114938 to the release.

17 November, 21:11 UTC. Calif published its technical write-up, “Write once, RCE everywhere”, with a demo video, the full exploitation technique, and a self-contained safe trigger. Its product-specific exploits stay private until users have had a chance to upgrade.

17 November. CISA issued its alert, “Critical Java Vulnerability Requires Immediate Patching”, with the Canadian Centre for Cyber Security contributing (alert AL26-029, classifying the flaw as a heap-based buffer overflow and reporting no awareness of active exploitation). Red Hat published RHSB-2026-018. Amazon released Corretto 17.0.21.1.2. Wiz said its Atlas AI harness built a proof-of-concept exploit within two hours, before the full details were disclosed, and confirmed remote code execution against Jira, Confluence and Jenkins. Tenable, Sophos and Qualys each reported no known exploitation; Cisco Talos reported none seen.

Sources: Calif; CISA alert; OpenJDK advisory; Oracle November CPU via Censys, Tenable and The Hacker News; Wiz; Red Hat bulletin RHSB-2026-018; Canadian Centre AL26-029 via The Hacker News and CyberScoop reporting.

↑ Back to contents

The vulnerability

What the bug is

Base64 is how binary data is commonly represented as text — in authentication protocols, file uploads and network messages. To decode it faster, OpenJDK’s HotSpot JVM substitutes hand-written assembly for the Java decoding loop once a method runs hot. The x86-64 Base64 intrinsic processes input in 256-byte blocks (64-byte blocks below that), using AVX-512 VBMI instructions, and handles the final few characters — the padded tail — in a separate code path. That path is where the bug lives.

For the 65-character input above, OpenJDK allocates a 50-byte result array: the decoded length is calculated as (ceil(input_length / 4) * 3) - padding. After decoding the first 64 characters, the routine mishandles the stray = and calculates a negative output length. Calif’s pseudocode of the faulty logic:

output_len := (input_length / 4) * 3
// handle padding
if input[input_length - 1] == '=':
    output_len -= 1
    input_mask >>= 1
    if input[input_length - 2] == '=':
        output_len -= 1
        input_mask >>= 1
output_mask := bzhiq(-1, output_len)

When the remaining input is of the form =, ==, X=, X==, XY= or ===, output_len is computed as zero and then decremented, underflowing to -1 or -2. Because bzhiq(-1, output_len) does not change any bits when output_len >= 64, the result is output_mask == -1u64 — all 64 bits set. The following evmovdqub (masked vector store) therefore writes all 64 bytes, starting at offset 48 of the 50-byte array: 62 bytes past its end (63 for a == tail). The generated stub returns early when it detects invalid input, and the Java code then throws IllegalArgumentException — but the write has already happened.

The write itself is almost entirely fixed: the attacker fully controls byte 0, partly controls byte 1 (4 bits), and bytes 48–63 are always 0x9a. Only the landing position is controllable. Calif’s exploitation strategy turns that fixed write into control by overwriting a buffer’s length — a classic JavaScript-engine primitive. The length-corrupted buffer becomes the writer: it keeps accepting chosen bytes, so a neighbouring array or object can be corrupted with chosen values. Corrupting a second array’s length, and having its data copied back into the response, builds the reader — a heap-disclosure primitive Calif calls JavaBleed, able to leak heap contents that may hold PII, private keys, other secret keys and tokens. From writer and reader, the chain repairs the objects the linear write damaged, mounts data-only attacks on application logic, and escalates the bounded write into arbitrary writes reaching the JVM’s code cache and hooks for code execution.

Getting there needs heap grooming. G1 divides the heap into fixed-size regions, and HotSpot normally gives each thread a private thread-local allocation buffer (TLAB), so objects created by different request threads have no useful order. Calif sprays several live writer/reader pairs, controls the allocation route (array size and remaining TLAB space decide whether HotSpot refills the TLAB or takes the attacker-preferred outside-TLAB path), and paces requests so each chunk is TCP-acknowledged before the next is sent, giving control over allocations up to the critical point. The attack is sensitive to network conditions — low, stable round-trip times are preferred — and Calif rented a VM in the same region as its target rather than tuning TCP/IP settings. The chain runs in stages: warm (valid decodes until the JIT selects the affected fast path), rehearse (an optional safe round for calibration), trigger (the malformed tail), disclose (enlarge the reader, collect response-derived heap bytes), repair (restore temporary headers and untouched neighbours), finish (make the application-specific change and verify it through a live request). A misplaced trigger can crash the JVM, but in many cases the attack can simply be retried.

Two limitations bound the attack. Some grooming profiles need partial request bodies to reach the application, and a proxy that buffers the complete body removes their release timing — nginx buffers request bodies by default — though that is not a general mitigation: Elasticsearch runs from one complete request, and after a WebSocket upgrade later frames pass through a tunnel where buffering no longer applies. And there is no universal payload: each chain needs a decoder route, a live writer, a returned reader or another feedback path, a heap layout and an application action.

Provenance of the code. The AVX-512 Base64 optimisation was written as JDK-8268276 for JDK 18 and backported to Java 17 as JDK-8296735 in the 17.0.6 update — which is why 17.0.6 is the boundary and why the vulnerable code has been shipping since January 2023. Endor Labs describes the fix as four lines. The trigger, in full:

AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA=

On older OpenJDK 17 builds the trigger is 128 A characters followed by =.

↑ Back to contents

Exploitation status

Demonstrated by the discoverers, disputed in scope, not seen in the wild

The central question of this case is not whether the overflow is real — every lab that tried reproduced it — but how far it reaches. The table below keeps each party’s claim separate from what anyone else confirmed.

PartyClaim / findingStatus
CalifPre-authentication RCE against Keycloak, Jira, Confluence, Artifactory, Jenkins and PingFederate in default configurations; post-authentication RCE against Kafka Broker, Solr, Elasticsearch, Gerrit and Spinnaker; RCE in Struts and Spring Boot samples; client-side exploit over TLS (samczsun)Medium — the discoverer’s own testing; not independently confirmed
WizAtlas AI harness built a PoC within two hours, before full disclosure; proved remote exploitation feasible against Jenkins, Confluence and Apache Tomcat; generalising to an arbitrary target is difficultMedium — second-party confirmation for three applications
Blackpoint Cyber (APG)Reproduced JVM crashes under normal runtime configurations, including repeated unauthenticated HTTP requests against a local Spring Boot application with Spring Security; reproduced the writer/reader heap disclosure locally; did not reproduce remote RCEHigh — independent lab, stated limits
Harness (Java2Shell)Reproduced the 62-byte overflow and a JVM crash with small fixed TLABs (SIGSEGV in the G1 collector); the application answers with an ordinary 401 under default heap settings; did not pursue code executionHigh — independent lab, stated limits
UpwindReproduced the overflow on an AWS c6i instance (Intel Ice Lake); -XX:-UseBASE64Intrinsics stopped it on every affected version testedHigh — independent lab
ImmersiveBuilt its own working exploitMedium — reported second-hand, no detail published
EquixlyRCE unproven; a JVM crash is the realistic outcomeMedium — dissenting assessment, not a reproduction
GreyNoiseProbing requests detected; no confirmed successful attackHigh — sensor telemetry
CISA, Tenable, Sophos, Qualys, Cisco TalosNo reports of exploitation in the wild as of 17 November; Talos expects probing to increase after disclosureHigh — as of 17–18 November
CISA KEVNot listed (catalogue v2026.11.17, 1,782 entries); CISA SSVC does not mark exploitation as activeHigh

Two readings of the same code coexist. Calif and Wiz describe a hard but demonstrated path from heap corruption to code execution, per application, with no universal exploit. Equixly and Harness read the same fault as pre-authentication heap corruption with a demonstrated denial of service, and treat code execution as unproven. Oracle’s own scoring sides with the harder reading: CVSS 3.1 8.1 with high attack complexity, described as “difficult to exploit”. What no party disputes: the overflow is reachable before authentication on any service that decodes attacker-controlled Base64 with the fast intrinsic, and the only complete fix is the patched runtime.

↑ Back to contents

Scope

Affected and fixed versions

Three conditions must all hold. Red Hat’s bulletin RHSB-2026-018 states them plainly: the runtime is OpenJDK 17.0.6 or later; the processor is x86-64 with AVX-512 VBMI; and the application passes Base64 data that an attacker controls to java.util.Base64.Decoder, without the JVM having been started with AVX-512 disabled. No special configuration is needed — the fast path is chosen automatically on capable CPUs, so affected systems are vulnerable out of the box.

Java lineAffected Oracle builds (and likely earlier)Fixed Oracle build (17 Nov)
Java SE 8 Enterprise Performance Pack8u511b33-perf8u513-perf
Java SE 17 / OpenJDK 1717.0.6 onward, through 17.0.21.0.317.0.21.1
Java SE 21 / OpenJDK 21Through 21.0.13.0.321.0.13.1
Java SE 25 / OpenJDK 25Through 25.0.5.0.325.0.5.1
Java SE 27 / OpenJDK 27Through 27.0.1.0.327.0.1.1
Oracle GraalVM for JDK17.0.21.0.3, 21.0.13.0.3Per Oracle’s advisory — fixed build not confirmed in the sources read
Oracle GraalVM Enterprise Edition21.3.20.0.3Per Oracle’s advisory — fixed build not confirmed in the sources read

Not affected. Standard Java SE 8 and Java 11 builds do not contain the vulnerable code path. ARM64 systems do not run it. x86-64 machines without AVX-512 VBMI do not run it. Upwind’s lab testing supports the boundary: affected were Temurin 17.0.6, 17.0.21, 21.0.13 and 25.0.5; clean were Java 11.0.33 and 17.0.5.

End-of-life lines. Java 18–20, 22–24 and 26 will not receive a fix; move them to a fixed LTS build. (IONIX, 18 Nov.)

Hardware. Intel Xeon Ice Lake, Sapphire Rapids, Emerald Rapids and Granite Rapids; AMD EPYC and Ryzen Zen 4 and newer (Red Hat also names Zen 5 and Zen 6). On Linux, check whether a host runs the vulnerable path:

grep -m1 '^flags' /proc/cpuinfo | tr ' ' '\n' | grep -E '^(avx512f|avx512bw|avx512vbmi|bmi2)$'

All four flags must be present. Oracle’s list covers only the builds it supports and says earlier versions are likely also affected; do not read Oracle’s build numbers onto other distributions — confirm the fix from your vendor’s own release notes.

↑ Back to contents

Patch rollout

Vendor patch status, 20 November

Oracle’s November Critical Patch Update covers 655 vulnerabilities across its portfolio, 62 of them rated critical. Eleven fixes are in Java SE, all described as remotely exploitable without authentication, with a maximum CVSS 3.1 score of 8.1. Three of the other ten (CVE-2026-114941, CVE-2026-114942 and CVE-2026-114945) affect only the Java 21, 25 and 27 builds Oracle lists. Oracle’s advisory does not publish CVE identifiers for the Java fixes, and its release notes for 17.0.21.1 do not name CVE-2026-114938 — the OpenJDK advisory is the document that ties the CVE to the release.

DistributionStatus, 20 NovFixed build
Oracle Java SEShipped 17 Nov17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1, 8u513-perf; 11.0.33.1 for other flaws in the same update
Amazon CorrettoShipped 17 Nov17.0.21.1.2, 21.0.13.1.2, 25.0.5.1.2, 27.0.1.1.4 — CVE named in the changelogs
Azul ZuluShipped17.70.201, 21.54.201, 25.38.201, 27.30.201 (free builds)
SAP SapMachineShipped, unconfirmed17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1 released this week, but the release pages do not say what they fix
Red Hat build of OpenJDK / RHEL 8, 9, 10ShippedBuild of OpenJDK 17.0.21.1, 21.0.13.1, 25.0.5.1 shipped 18 Nov (RHSA-2026:100565, 100566, 100569, 100570, 100573); RHEL RPMs shipped 19 Nov — java-17 17.0.21.1.2 for RHEL 8 (RHSA-2026:100563) and RHEL 9 (RHSA-2026:100564), java-21 21.0.13.1.2 for RHEL 8 and 9 (RHSA-2026:100567) and RHEL 10 (RHSA-2026:100568), java-25 25.0.5.1.2 for RHEL 9 (RHSA-2026:100571) and RHEL 10 (RHSA-2026:100572). Red Hat container images and the Red Hat Build of Keycloak remain affected
Eclipse TemurinShipping — x64 Linux + macOS 17 done17.0.21.1+2, 21.0.13.1+2, 25.0.5.1+2, 27.0.1.1+4 for x64 Linux published 18 Nov; Java 17 build for macOS x64 published 19 Nov; Windows builds and Java 21/25 macOS builds still in testing
Microsoft Build of OpenJDKShipped 18 Nov17.0.21.1, 21.0.13.1, 25.0.5.1 — CVE named in the release notes
BellSoft LibericaShipped17.0.21.1+6, 21.0.13.1+6, 25.0.5.1+5, 27.0.1.1+5 published 18 Nov; BellSoft blog (18 Nov, updated 19 Nov) confirms CSPU builds 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1 include the patch for CVE-2026-114938
Okta (Access Gateway, LDAP, provisioning and RADIUS agents)Shipped 19 NovOkta Access Gateway 2026.45.1 or greater; LDAP Agent 5.26.3 or greater; On-Premises Provisioning Agent 3.3.4 or greater; RADIUS Agent 2.30.0 or greater. Patches rolling out across Okta production environments; Auth0 not affected

Until a vendor’s fixed build ships, the JVM flags in the action plan are the mitigation. Berkeley’s Information Security Office, citing Oracle and the OpenJDK advisory, lists the same Oracle replacement builds; Oracle advises moving every line, including 8 and 11, to the November releases, because the update fixes other flaws beyond this CVE.

↑ Back to contents

Demonstrated impact

Applications Calif says it exploited

These are Calif’s results from its own testing, in the applications’ default configurations. They have not been independently confirmed, and Calif stresses the lists are non-exhaustive: any endpoint that Base64-decodes attacker-controlled bytes is a candidate. Wiz proved remote exploitation feasible against Jenkins, Confluence and Apache Tomcat; its proven-vulnerable list also names Jira, Keycloak and Artifactory, the post-authentication five below, and the Spring Boot and Struts frameworks. Apache Tomcat appears in Wiz’s list, not in Calif’s.

Pre-authentication RCE (Calif)What it is
KeycloakIdentity and access management; single sign-on for staff and customers
Atlassian JiraIssue and project tracking, widely deployed in engineering and operations
Atlassian ConfluenceCollaboration wiki, commonly internet-facing or widely reachable inside firms
JFrog ArtifactoryArtifact repository; build and release infrastructure
JenkinsCI/CD automation; a foothold here reaches build and deployment pipelines
PingFederateEnterprise identity federation and single sign-on
Post-authentication RCE (Calif)What it is
Apache Kafka BrokerEvent streaming; the backbone of many payment and messaging pipelines
Apache SolrSearch platform
ElasticsearchSearch and analytics; often holds log and customer data
GerritCode review and repository management
Netflix SpinnakerContinuous delivery platform

Calif also demonstrated RCE in sample applications built on Apache Struts and Spring Boot, and reports a client-side case: researcher samczsun shared an exploit in which a Java program that opens a TLS connection to a malicious server can be compromised, even if it exposes no service of its own. Ordinary features such as authentication and file uploads can expose the vulnerable decoder — Base64 decoding runs almost anywhere a Java service parses input: authentication tokens, file uploads, network messages. Harness notes that HTTP Basic credentials, JWT segments and SASL tokens all pass through java.util.Base64 while the caller is still anonymous.

↑ Back to contents

Exploit availability

What exploit code exists, and what is withheld

Calif’s safe trigger. Calif published a self-contained proof of concept in its write-up. It warms the Base64 decoder until the vulnerable intrinsic is used, then decodes the malformed 64 × “A” + “=” input into a 4096-byte array. On an affected JDK it prints REPRODUCED: the decoder rejects the input, but 62 bytes (offsets 50 to 111) past the decoded output have already been overwritten. Every write lands inside a real array, so the trigger is safe to run. The product-specific exploits — the weaponised chains for individual products — stay private until everyone has had a chance to upgrade; Calif says it will publish them later.

GitHub. cve.now counts 55 proofs of concept on GitHub as of 18 November, the first appearing on 18 November at 02:15 UTC (Ghost-Bx1/java-exploits, r3dvei1/java4shell). The count rose from 24 during 18 November; none of the newer repositories has been verified. One is linked here: arkrwn/PoC-CVE-2026-114938-CrashLeak (created 18 November 03:37 UTC, single commit). Its README and its full Rust code were read before linking: a crash and heap-leak reproducer — a bounded local lab oracle plus a network trigger harness — not a universal RCE shell, as its author states. It includes optional command-proof flags (--whoami, --listen) that prove command execution only on an endpoint that is already injectable, not the withheld Java4Shell chain. The second is ikow/java4shell-poc (created 17 November 23:04 UTC; author Qi Deng, writeup on Substack; 131 stars and 48 forks as of 18 November) — the most complete public toolkit so far: the bounded trigger, a five-stage chain (writer, JavaBleed disclosure, heap feng shui, full chain, remote simulation), per-target modules for Keycloak, Jenkins, Spring Boot and Elasticsearch, host and network scanners, and Suricata, Sigma and ModSecurity detection rules. Its author reports remote heap corruption confirmed against Keycloak, Jenkins and Elasticsearch in default configurations — heap corruption, not RCE; the product-specific RCE chains remain Calif’s withheld work. The author’s own writeup cautions that confirming the overflow is straightforward and exploiting it is not. Its README, writeup and the key code files were read before linking: the scanner and exploit code talks only to a target the user supplies, with no exfiltration host and no unexpected network calls. The other repositories have not been verified individually and are not linked; fake PoC repositories delivering malware are a known pattern after high-profile disclosures, so read any script before running it. Calif’s write-up, which contains the trigger inline, remains the verified source for the exploit chain itself.

Detection tooling. Harness published a runtime checker script (linked from its blog) that compares each local JDK against the October 2026 advisory builds to report whether a host is vulnerable. Wiz customers can use a pre-built query and advisory in the Wiz Threat Center to search for vulnerable instances. Endor Labs describes the fix itself as four lines.

↑ Back to contents

Detection and hunting

Hunt the effect, not the request

No network indicators have been published. There are no IPs, domains, file hashes or malware families tied to Java4Shell exploitation, because no exploitation in the wild has been reported. Do not import indicators from unrelated Java or OpenJDK incidents. What exists to hunt on is behaviour:

↑ Back to contents

Action plan

What to do now

  1. Inventory the runtime, not the application. Review software inventories for JDKs and Java runtimes, including those bundled with third-party applications or included in container images and appliances. CISA warns these may not appear as separate entries in software inventories, and Sonatype notes that project dependency scans will not identify every affected installation — the flaw is in the runtime, not a library. Check the version each running service reports, not just what the package manager says.
  2. Patch the runtime; patching the application does not fix it. Install your vendor’s fixed build from the table above and restart the JVMs. Prioritise internet-facing systems first (CISA). Oracle’s fixed builds are 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1 and 8u513-perf; Corretto, Azul, Microsoft and Temurin (x64 Linux) have shipped, Red Hat has shipped fixed builds for its Build of OpenJDK and all six RHEL errata, BellSoft has confirmed fixed CSPU builds, and Okta has published fixed agent versions; the Red Hat container images and the Red Hat Build of Keycloak remain affected — track that vendor’s advisory rather than waiting for a generic alert.
  3. Where patching must wait, disable the intrinsic. Start the JVM with -XX:-UseBASE64Intrinsics to turn off the hardware-accelerated Base64 routines and fall back to the software decoder. This is the recommended flag: it works on every architecture. The alternative, -XX:UseAVX=2, disables AVX-512 — but OpenJDK on ARM does not recognise it and will fail to start, so applying it blindly across a mixed estate will crash ARM hosts. Apply -XX:UseAVX=2 only on x86-64 hosts. Both flags may carry a performance cost.
  4. Compensating controls. Where immediate patching is not feasible, CISA recommends restricting and validating Base64 input at the application boundary, enforcing input-size limits, and limiting unauthenticated access. Wiz recommends temporarily restricting external network access to vulnerable management portals (Jira, Jenkins, Keycloak) until mitigations or updates are deployed.
  5. Do not rely on the network edge. WAF rules and request-body buffering (nginx buffers bodies by default) blunt some attack profiles, not all: Elasticsearch runs from one complete request, and WebSocket frames pass through a tunnel after upgrade. Treat edge controls as delay, not defence.

↑ Back to contents

BFSI relevance

Why banks should care

No bank has been named as a victim, and no financial-sector alert has been issued. That is the whole of the victim record so far, and it should stay that way unless a named institution discloses. The relevance is structural, not incident-driven — and it is large.

Java runs the sector. Calif puts it bluntly in its write-up: Java powers 90% of Fortune 500 companies, and “banking, aviation, telecommunications, nearly every critical sector runs on Java, including national security systems.” Core banking systems, payment switches, trading platforms, messaging and event-streaming backbones (Kafka is on Calif’s post-authentication list) and the identity layer all commonly run on the JVM. Two of the six products Calif says it took over without logging in are identity systems — Keycloak and PingFederate — the single sign-on that guards staff access to everything else. Jenkins, Jira and Confluence are standard tooling inside financial IT. A runtime flaw needs no vulnerable application: any Java service that decodes attacker-controlled Base64 on affected hardware is a candidate, and Base64 is how authentication tokens, file uploads and network messages are commonly represented.

The practical exposure for a bank is the same inventory trap CISA describes, sharpened by the sector’s vendor estate: Java runtimes bundled inside third-party products, core-banking appliances and container images, on patch cycles the bank does not control, with the vendor’s fixed build as the only route. Red Hat’s bulletin naming Azure D-series instance families shows the exposure extends into cloud estates banks rent rather than own. Patch the runtime you run, chase the runtime your vendors run, and treat the two JVM flags as the bridge where a vendor’s build has not shipped — Red Hat’s remaining RHEL packages are the ones that matter most on 19 November, now that its Build of OpenJDK, Temurin (x64 Linux) and Microsoft have shipped.

Covered in the daily brief of 18 November 2026 (11182026.html) as the edition’s lead vulnerability; this page is the deep-research record for it.

↑ Back to contents

Exposure

How much Java is actually out there

Unknown, and that is the honest answer. No vendor or sensor operator has published a count of internet-facing hosts running a vulnerable OpenJDK build on AVX-512 VBMI hardware, and none is quoted here. What is known:

Exposure is therefore best measured per estate — how many of your JVMs run 17.0.6 or later on AVX-512 VBMI hardware, including the ones inside vendor products — rather than from any global count. If a global count appears, it belongs here with its methodology, not as a bare number.

↑ Back to contents

Timeline

How a 2023 optimisation became a November emergency

↑ Back to contents

Confidence calibration

What is settled, and what is not

The standing rule for this project is to state what is missing rather than round up. On day two, root cause, affected range and fixed builds are settled; the reach of exploitation is not.

ClaimConfidenceWhy
The overflow exists, in the AVX-512 Base64 intrinsic, with the stated triggerHighCalif’s write-up with pseudocode and trigger; OpenJDK and Oracle confirm the component; Blackpoint, Harness and Upwind reproduced the overflow independently
Affected range: OpenJDK 17.0.6 onward, lines 17/21/25/27; standard 8 and 11 clean; ARM64 cleanHighCalif, OpenJDK risk matrix, Oracle’s build list, Red Hat bulletin, Upwind lab testing
Oracle fixed builds shipped 17 NovHighOracle November CPU; release notes for 17.0.21.1, 21.0.13.1, 25.0.5.1, 27.0.1.1
Pre-authentication RCE against the six named productsMediumCalif’s own testing, plus Wiz for Jenkins, Confluence and Apache Tomcat; no independent reproduction of full remote RCE; Equixly and Harness dispute the reach
Score 9.2 under CVSS 4.0MediumAttributed to Calif by most reporting; at least one secondary source attributes it to South Korea’s NCSC; Horizon3 notes the source of the 9.2 is not named in the CVE record; no CVSS 4.0 vector published
Score 8.1 under CVSS 3.1, high attack complexityHighOracle’s CVE record, the OpenJDK risk matrix and the published NVD record; vector published (AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H)
Red Hat 9.8MediumRed Hat’s own CVE page now publishes the vector (CVSS:3.0/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H); attack complexity low is the whole disagreement with Oracle’s high; still described as preliminary in reporting; fixed builds for the Build of OpenJDK shipped 18 Nov (RHSA-2026:100565/100566/100569/100570/100573); all six RHEL errata shipped 19 Nov (RHSA-2026:100563, 100564, 100567, 100568, 100571, 100572)
No exploitation in the wild as of 20 Nov; not in CISA KEVHighCISA, Tenable, Sophos, Qualys, Cisco Talos, GreyNoise; catalogue v2026.11.18, 1,786 entries
Details leaked before coordinated disclosureLowCalif’s claim only; no detail on how, and no independent confirmation
GraalVM fixed buildsLowOracle lists the affected GraalVM builds; no fixed build confirmed in the sources read — check Oracle’s advisory

↑ Back to contents

Lessons learnt

What this case shows so far

Two days in, the record already returns a short list. Each lesson below is drawn from the record above; where the record is still open, it stays open here.

The runtime is the blast radius, not the application

Nothing in Jira, Jenkins or Keycloak was at fault. The flaw sits in the Java runtime every one of them shares, which is why a single bug credibly threatens six unrelated products at once — and why patching the application cannot fix it. Inventory and patch the runtime; the application list is a triage aid, not the fix.

AI has compressed discovery-to-exploit to hours

One researcher, one hour to find the bug, two more to a first working exploit, under $200 in token costs; a second firm’s harness built a PoC in two hours before the details were public. The patch window is no longer measured against human exploit development. The same compression applies to defenders only if their inventory already knows which JVMs they run.

Score disagreement is a signal, not noise

9.2, 8.1 and 9.8 describe the same flaw. The gap is not arithmetic — it is a disagreement about exploitability, between a researcher who built the chain and a vendor that calls it difficult. Read the scorer and the attack-complexity judgement, not the number; and note when a score’s source is unnamed and its vector unpublished.

Bundled runtimes are the inventory gap

CISA’s warning is the operational core of this case: runtimes inside third-party products, containers and appliances may not appear in software inventories, and dependency scans cannot see a runtime flaw at all. The estate you can list is not the estate you run.

“No universal exploit” cuts both ways

Calif’s caveat is genuine — each target needs its own chain — but it is not reassurance. It means there is no single signature to detect and no single patch to the application layer; it also means a motivated adversary can tailor the chain per target, with AI, for the products that matter most. The absence of a generic exploit is not the absence of risk.

Withheld exploits buy time only if patching uses it

Calif is holding its product chains back until users have had a chance to upgrade, and other labs have so far reproduced crashes rather than shells. That window is measured in days: probing has already begun, GitHub PoCs are appearing, and a second firm says it has built its own working exploit. The value of the delay is spent on patching, not on reassurance.

↑ Back to contents

Source register

Everything this brief rests on

Source register — click to collapse

↑ Back to contents

← Back to main page