TLP:AMBER Living tracker · first published 22 Oct 2026 · repurposed 9 Nov 2026

Tracking offensive AI tooling in the wild

A living tracker of agentic offensive-security frameworks — ARTEX, CyberStrikeAI, PentAGI, Cairn and HexStrike — as they are used in real intrusions, with attack paths, TTPs and IOCs per tool. The ConnectWise breach review this page started as is kept below as a case file.

Sources: 110 listed Indicators reviewed: see Detection Queries run: 22 Oct 2026; re-checked to 20 Nov
The shift HIGH

Agentic offensive frameworks — built as legitimate pentest tools — are now doing live intrusions for criminals and state-aligned actors alike, at machine speed and for dollars per target.

Confirmed in vendor incident response (GTIG, Microsoft, Kaspersky, TrendAI), government advisory (CERT-In CIAD-2026-0058) and operator infrastructure exposed in the open.

Attribution MEDIUM

Five frameworks, many operators. The tool name is the weakest link: shared tooling does not mean a shared actor.

Operators stack frameworks on the same hosts. Hunt the operator’s behaviour — infrastructure, targets, model, post-exploitation tooling — not one tool’s fingerprint.

1,465exposed ARTEX consoles (Censys, 4 Nov)
3,000+CyberStrikeAI instances in use (ThreatBook, vendor claim)
600,000+payment-card records stolen in the Cairn retail campaign (Gambit, Sep 2026)
459exposed PentAGI consoles (RST, 2 Nov)
The finding that changes your hunt: operators stack these frameworks — Censys found CyberStrikeAI on 70 ARTEX hosts, and one host ran ARTEX and PentAGI beside a Cobalt Strike teamserver and an Adaptix C2 listener. A detection that names one tool and stops there will miss the operator.

The tools at a glance

Five frameworks, all open source, all built as legitimate pentest tooling, all now observed in live intrusions. Status is what the sources establish, not what the tool could do.

ToolOriginWhat it doesConfirmed nefarious useBFSI relevance
ARTEXAutumn-27, Jul 2026, AGPL-3.0, Go / Next.jsMulti-agent autonomous pentest: recon to exploitation to collection, model-agnosticStorm-0940 / ConnectWise, Storm-3358, VAPOR PANDA, Crafty Falcon; 1,465 exposed consolesDirect: Korean finance, Indian BFSI payment APIs, SE Asia finance
CyberStrikeAIEd1s0nZ, Nov 2025, Apache-2.0, GoLLM-orchestrated pentest, 100+ tools, MCP, multi-agent, built-in C2Kaspersky education intrusions (domain admin in 3–4 days); TrendAI 44-agent Japan swarm; Crafty FalconDirect: CERT-In names it in BFSI payment-API theft
PentAGIVXControl, Jan 2025, MITMulti-agent pentest, 20+ Kali tools in sandboxes, 10+ LLM providersGTIG / Mandiant IR cases; Crafty Falcon; CyberXero initial-access broker; 459 exposed consolesDirect: CERT-In BFSI campaign; exposed consoles are hijackable
Cairnoritera / Cairn, Apr 2026, AGPL-3.0, PythonBlackboard-coordinated agent harness; coding-agent workers, no auth on its console by defaultGambit retail card theft (600,000+ cards); NorForge / VULONE campaigns; Crafty FalconDirect: payment-card theft, payment APIs, exposed panels leak credentials
HexStrike AI0x4m4, Aug 2025, MITMCP server, 150+ tools, 12+ agents, natural-language to exploit chainNetScaler zero-day abuse (reported); BOSSAI mesh (confirmed component)Direct: NetScaler edge devices, lending-firm breach; the tool itself has CVE-2025-35028 (9.1)

Exposure counts are console counts, not malicious-operator counts: many exposed consoles belong to legitimate testers. Treat a count as the size of the discoverable surface, not as a victim or attacker count. Each tool section below carries its as-of date.

ARTEX

Open-source, LLM-driven autonomous penetration-testing framework · developer known as Autumn-27 · published on GitHub as Autumn-27/ARTEX on 26 July 2026 under AGPL-3.0 · Go backend, Next.js web console, PostgreSQL · upstream repository taken down / made closed-source by the developer in October 2026 after abuse reports; forks and copies continue to circulate

A multi-agent framework that autonomously runs a penetration test from reconnaissance through exploitation, privilege escalation and data collection, with little human input: it builds an asset graph of the target and an exploration graph of what has been tested, discovers exposed services, probes for vulnerabilities, generates and refines attack scripts, exploits them, and records findings and credentials. Model-agnostic — reported use of DeepSeek, Kimi, GLM and Grok via resellers and shims. Default console port 8787.

Timeline of reporting

All times UTC unless dated otherwise. Corroborated by at least two independent observers unless marked as a vendor assertion. The ConnectWise compromise is one episode in this timeline, not a separate case. Newest first.

Entries are tagged by the tools they name. Entries naming no tool — or stating explicitly that no tool is linked — appear under All only.

Attack path — new 9 November finding

Single source: H3NRY B41T on Medium, 9 Nov 2026, based on Hunt.io AttackCapture data. An open directory (1,235 files, 309 MB, 566 flagged malicious) on a staging server exposed an ARTEX-assisted campaign running 2–6 Nov against enterprise infrastructure in Japan (primary), Taiwan, Korea and the US. Status: reported — no vendor or government corroboration as of 9 Nov, and no actor name or state link is asserted.

Gap: this source does not describe a data-exfiltration step for this operation — no exfiltration path, destination or volume is given, so that stage is unknown for this campaign. The generic ARTEX path described by Microsoft and CrowdStrike in October (recon, password spraying / OAuth abuse, privileged cloud access, exfiltration) is background, not a finding from this window. Two compromises are claimed in the source: a captured phpMyAdmin admin session cookie for 107.167.177.120, and a CVE-2023-46604 RCE callback log showing a Taiwanese target fetching the operator’s Spring XML payload. Treat both as reported, not independently confirmed.

TTPs

MITRE IDs as published in the source’s own ATT&CK matrix (Medium, 9 Nov 2026). No mapping was invented.

TechniqueMITRE ID
Reconnaissance — Active Scanning (FOFA API queries, masscan logs, probe scripts)T1595
Reconnaissance — Gather Victim Network Information (BI platform enumeration: Metabase, Superset, Airflow)T1590
Reconnaissance — System Language Discovery (locale detection in scripts)T1614.001
Initial Access — Exploit Public-Facing Application (CVE-2023-46604, Advantech WebAccess, F5 BIG-IP, Struts2, GlobalProtect)T1190
Execution — Command and Scripting Interpreter: PythonT1059.006
Execution — Windows Command ShellT1059.003
Persistence — Server Software Component: Web Shell (phpMyAdmin SQL OUTFILE webshell attempts)T1505.003
Persistence — Exploitation for Persistence (Redis rogue master payload delivery)T1190
Defense Evasion — Virtualization / Sandbox Evasion (system checks in scripts)T1497.001
Defense Evasion — Data Staged: Local Staging (/dev/shm memory-only staging)T1074.001
Credential Access — Credentials from Password Stores (WinboxExploit.py extracting user.dat)T1555
Credential Access — Steal Web Session Cookie (phpMyAdmin session cookie in c2.txt)T1539
Credential Access — Brute Force (capbrute.py RuoYi math captcha bypass)T1110
Discovery — System Information Discovery (fingerprinting scripts)T1082
Discovery — System Network Configuration Discovery (network enumeration)T1016
Collection — Automated Collection (batch_run.sh, bi_stage1.py automated scanning)T1119
Command and Control — Application Layer Protocol: Web (cb_server.py, oob.py HTTP callback listeners)T1071.001

IOCs — new 9 November

From the Medium / Hunt.io report, 9 Nov 2026. File-hash IOCs are as published in that article; they were not independently verified.

ba8b0689388a2c07320d2257335f56d2690be76bff8a82a78d9e1a2db1fd3e16  artex_optimize.sh (2026-11-06)
ed419620e3e3a10a4360313892bd9a39e84e8f4a47957dc97c6167a705816955  amq_proof.py (2026-11-05)
a57dc9236beaec48e8a1bbc2128654afa9994a62e53a79cb418b75c14f5c800c  bi_stage1.py (2026-11-04)
8ba55951f0b84ba00b1d976df555f8f4d765dc6d5c7cd4205dbdca83abf9bd11  capbrute.py (2026-11-04)
84df1bd248b12d40d9ed83642aa0b537749f754d3aa82ad054883fe5683d379e  cb_server.py (2026-11-05)
c6dcf8f253f49af2358b87b40148f1e3a04740589818872fb779173f8c2f6964  c2.txt (2026-11-05)
4cedc5082bc05b7981311e90306873fe8fc3e8ce384438f03389502205be58ad  amq1555_tw_cb.log (2026-11-05)
bd5ef9be5533f55ab9d6cc88d93125709d7c871314db9861407e340c9e9a0845  act_jp.jsonl (2026-11-06)
efb105bde4d5087aa8423ca0ee1950e6c7cd64c0a8d38224061df1565aef0bd6  act_tw.jsonl (2026-11-06)
6918a56f0e2e7880261a04afc19d1d13b8479327e305e3e96f985ade750007da  act_us.jsonl (2026-11-06)

Exposure counts, two methods: GitGuardian, published 9 Nov 2026 from a 7 Nov Shodan scan, counts 1,428 ARTEX-family services; Censys, published 5 Nov, counts 1,465 exposed ARTEX consoles as of 4 Nov (~815 distinct-looking deployments after excluding ~650 hosts in two identically configured hosting blocks). The counts are measured differently and are close, not contradictory — and neither is a malicious-operator count. No named victim or actor attribution appeared to 12 November; Salesforce reported ARTEX attempts against customer environments on 12 November — a new target class (SaaS / CRM), with no IOCs published.

CyberStrikeAI

Open-source, Go-based, AI-native autonomous penetration-testing platform · developer alias Ed1s0nZ · first published on GitHub as Ed1s0nZ/CyberStrikeAI on 8 November 2025, Apache-2.0 · acquired by ThreatBook on 30 September 2026, which says it will keep the open-source edition available and launch an Enterprise Edition

An LLM orchestrator on the control server plans actions, selects tools and generates commands (exposed to the model as MCP tools); a plain implant / beacon on the victim host (no AI in the implant) polls via HTTP/S, TCP or WebSocket and executes shell commands and file operations. It integrates 100+ security tools and tool templates (recon, scanning, exploitation, analysis, reporting), with YAML-defined attack recipes, native MCP support, multi-agent orchestration, visual attack-chain tracking, RAG knowledge retrieval, WebShell management and built-in C2, human-in-the-loop approval, and automated structured reporting. Supports Claude, DeepSeek and other OpenAI-compatible models as the reasoning core. Its own documentation warns it is high-privilege: it can execute commands, invoke MCP tools, manage web-shell connections and optionally support C2 workflows.

Timeline of reporting

Dated entries, each linked to its source. Status is what the source establishes.

Attack path

Two sourced paths, plus a bounded third. No source publishes a single end-to-end CyberStrikeAI attack path covering all campaigns.

Nefarious use

TTPs

No primary vendor or government source in this research published MITRE ATT&CK IDs for CyberStrikeAI. All entries below therefore carry no MITRE ID; secondary commentary adds generic mappings, which were not adopted. The behaviours are as published by Kaspersky, TrendAI and CERT-In.

TechniqueMITRE ID
Initial access via compromised Microsoft SQL Server 16.0.1000; first malicious commands originated from sqlservr.exeNo mapping published
PowerShell payload execution; commands additionally Base64-encoded to bypass keyword detection; payloads split across multiple files and decoded with certutilNo mapping published
Persistence via scheduled task named SystemUpdate (or HKCU Run key value of same name); on Linux via crontab @reboot or ~/.bashrc entry; persistence method chosen by the LLM, subject to human-in-the-loop approvalNo mapping published
Discovery: local / domain / remote recon via PowerShell, net.exe, LDAP queries, WMI checks; checks for SCCM / AlwaysInstallElevated, LAPS, Azure / .azcopy; external-IP check via dnslog.cnNo mapping published
Credential access: LDAP queries for Kerberoasting and AS-REP Roasting; registry-hive dumping via reg.exe; LSASS process dump via comsvcs.dll MiniDump (using PID, not name); DPAPI secret decryptionNo mapping published
DCSync and Golden Ticket attacks via renamed mimikatz after compromising built-in domain administratorNo mapping published
Lateral movement via Impacket WMIExec and SMB, using shared network admin folders and WMI; reverse-shell script (p1.ps1) distributed via WMI; scheduled tasks for SYSTEM executionNo mapping published
C2: implant polls listener via HTTP/S, TCP or WebSocket; file transfer and screenshots via main encrypted channel, data Base64-encoded in standard JSON responses; LLM orchestrator on C2 server plans, selects and generates commands via MCP toolsNo mapping published
Collection / exfiltration: LLM-generated SQL / LDAP queries tailored to the target, recursive file scans, database export, screenshots (Windows and Linux)No mapping published
Automated full kill-chain workflow across 44 agents (recon, vuln discovery / validation, exploitation, C2, persistence, exfiltration, automated reporting) and generation of Japanese-language forms / fake login pagesNo mapping published
Mass discovery with nmap / masscan and exploitation with Metasploit modules (FortiGate campaign workflow)No mapping published
Payment / disbursement API abuse: invoking victims’ own fund-transfer APIs from allow-listed infrastructure, micro-value test transactions, log deletion and endpoint-security tampering (Crafty Falcon campaign-level TTPs; not attributed to CyberStrikeAI alone)No mapping published

IOCs

Team Cymru IPs are hosts running CyberStrikeAI in the FortiGate window; Kaspersky hashes are the education-campaign beacons and tooling. The Kaspersky Network indicators section returned empty in the fetched text, so no C2 domain from that report is listed.

212.11.64.250  103.164.81.110  106.52.47.65   115.120.233.95  117.72.103.145
118.25.186.119  142.171.160.137 144.31.224.253  146.190.195.154 146.190.82.132
154.219.114.92  38.38.250.182   43.106.25.225   43.167.237.212  47.101.186.156
47.95.33.207    60.204.227.64  62.234.61.215   64.176.48.93    81.70.144.252
2400:d321:2308:1461::1
185.186.245.144:5501  (Kaspersky p1.ps1 reverse-shell server)
108.165.147.226:8299  (Censys: host also running FRP and CyberStrikeAI; open directory with beacon builds)
45.87.53.3, 45.87.53.4, 45.87.53.6, 45.87.53.10, 45.87.53.14  (Censys: five hosts each running ARTEX, CyberStrikeAI, ARL and Vshell)
D726A8B4886C636F76CA6D30EF6A6FF4  Windows beacon, c:\windows\temp\svchosts.exe
2770562d104391b92b42671c270079bb  Windows beacon, svcmon32.exe
09e1da2afa1216ecc55648ec8e122d2a  Linux beacon
805B9E4AE74C903787416800999B8418  Linux beacon
29EFD64DD3C7FE1E2B022B7AD73A1BA5  wdiag2.exe, mimikatz
74F57B3E897A4960DCDD5B7E8A54F9DF  RogueWinRM.exe
0BBD08AF95F877F1D568B03DE2EB76B4  p1.ps1 reverse-shell script

Exposure: 3,000+ instances in use worldwide, per public internet-scanning data cited by ThreatBook (vendor claim, no independent audit published), as of 30 Sep 2026. Other counts, differently defined: Team Cymru 21 unique IPs (20 Jan – 26 Feb 2026); Censys 70 hosts co-locating CyberStrikeAI with ARTEX (as of 3–4 Nov 2026). GitGuardian (9 Nov, scan 7 Nov) adds a first exposed-service count: 2 on a strict fingerprint, 19 by page title.

PentAGI

Self-hosted, fully autonomous multi-agent penetration-testing platform · developed by VXControl L.L.C-FZ, Dubai Meydan Free Zone · GitHub repository vxcontrol/pentagi, MIT licence · repository created 6 January 2025, first public alpha v0.1.0 on 7 January 2025 · paid Enterprise Edition exists alongside the free Community Edition

The operator describes a target and a goal in plain language; a primary agent delegates to specialised agents (research, coding, installation, pentesting, adviser / reflector roles) which split the work into ‘flows’, sub-tasks and actions, run 20+ professional tools — nmap, Metasploit, sqlmap and other Kali Linux tools — inside isolated Docker sandboxes, use a built-in browser and search engines (Tavily, Firecrawl, Perplexity, DuckDuckGo, Google, Sploitus, SearXNG, Traversaal) for fresh intelligence, store results in PostgreSQL + pgvector (optional Neo4j / Graphiti knowledge graph), and produce a vulnerability report (web, Markdown or PDF). It supports 10+ LLM providers (OpenAI, Anthropic, Gemini, AWS Bedrock, Ollama / local models and other OpenAI-compatible endpoints) via REST / GraphQL APIs. It is explicitly not a CALDERA-style breach-and-attack-simulation product with predefined campaigns; the agents decide the next step. Use is restricted by its EULA to systems the operator owns or is authorised to test.

Timeline of reporting

Dated entries, each linked to its source. Status is what the source establishes.

Attack path

No source publishes a complete, PentAGI-only kill chain, so the path below is assembled from campaign-level descriptions and the gaps are marked.

Gap: which steps in Crafty Falcon were executed by PentAGI versus ARTEX / Cairn / CyberStrikeAI is not stated in the public advisory; no PentAGI-specific persistence, credential-access or exfiltration technique was published.

Nefarious use

TTPs

Only one MITRE ID was published for PentAGI specifically (SOCRadar’s CyberXero table). The rest are behaviours as published, with no ID invented. The other MITRE IDs in SOCRadar’s CyberXero table describe the actor’s overall operation, not PentAGI, and are not listed here.

TechniqueMITRE ID
Software Deployment Tools — PentAGI orchestrating Cobalt Strike operations on the Team Server via API (payload generation and pentest execution)T1072
Active vulnerability scanning and WHOIS asset mapping (systematic reconnaissance and vulnerability scanning against organisations in Asia)No mapping published
Intrusion and persistence following scanning (technique unspecified in source)No mapping published
Use of publicly available AI offensive tooling to survey targets with machine intelligence, generate / refine large volumes of tailored attack scripts in rapid succession, adapt payloads to target responses, and rotate supporting IP addresses within hours to daysNo mapping published
Payment / disbursement API abuse from the victim’s own allow-listed infrastructure, preceded by a micro-value (Rs.1) test transfer to validate the fund flowNo mapping published
Entry via internet-exposed non-production (UAT / staging / test) environments, pivot to production via shared credentials / tokens / network paths; cloud management / serverless abuse (AWS Systems Manager, Lambda); IP allow-list bypass via whitelisted third-party fetcher poolsNo mapping published

IOCs

Exposure: 459 verified internet-exposed PentAGI consoles (unique IPs, deduplicated across Shodan and Netlas; default port 8443, 69% on default port; top hosting countries US 43%, Hong Kong 11%, China 11%, Germany 7%; about 90 consoles in one network, AS398823), RST Cloud, as of 2 Nov 2026. Count is a lower bound — renamed / rebranded builds and non-standard ports are missed by title searches. For comparison, Censys alone showed 506 PentAGI consoles on 2 Nov 2026. GitGuardian’s 7 Nov scan counts 419 by favicon hash (521 by page title) — again a different method, not a contradiction. Cisco Talos (13 Nov, Censys data) counts more than 700 currently reachable PentAGI instances — again a different method, not a contradiction. Exposure does not prove malicious use: many consoles will belong to legitimate testers. No PentAGI-specific IOCs were published in the public CERT-In CIAD-2026-0058 advisory; fingerprints / detection rules were shared only in the restricted TLP:AMBER version to BFSI entities.

Cairn

Open-source project on GitHub at oritera/Cairn · repository created 19 April 2026 · author posts as ‘le1xia0’ (organisation ‘oritera’) · licence AGPL-3.0 · written mainly in Python · latest release v0.2.1 (10 May 2026)

Name collision — do not conflate. A second, unrelated tool is called CAIRN: Cisco Talos’s CAIRN (Cognitive Artifact Intelligence Research Network), a defensive, metadata-only toolkit for hunting AI-integrated malware, released 22 September 2026. This entry covers oritera/Cairn, the open-source AI agent harness abused in live intrusions. Conflating the two would invert the entry.

A general-purpose problem-solving engine. The operator supplies an origin (starting point) and a goal; Cairn searches for a path through an unknown state space. Coding-agent workers (Claude Code, Codex and Pi CLIs) run observe-orient-decide-act loops and coordinate only through a shared ‘blackboard’ — a fact / intent / hint graph in which facts are confirmed findings, intents are next steps, and hints are human input. It defines no fixed roles or workflows. A dispatcher assigns tasks to workers (often in disposable Kali containers); a web console (default port 8000, served by uvicorn) shows projects, graphs and findings. Critically for defenders, Cairn ships with no authentication by default on that console — anyone who finds it can view projects, secrets and findings, and start, stop or redirect work. Built for, and first validated on, autonomous penetration testing: it solved 54/54 problems in the 2nd Tencent Cloud Hackathon AI penetration-testing challenge (610 teams, 1,345 participants), the only team to do so, finishing 3rd overall — per the project’s own README. The README states it is intended for use only where the operator has explicit authorisation.

Timeline of reporting

Dated entries, each linked to its source. Status is what the source establishes.

Attack path

No source describes one canonical end-to-end path for all Cairn use — the path varies by campaign, and the composite below is assembled from the separate cases; gaps are marked.

Gaps: exfiltration volumes are largely unconfirmed (Reuters could not ascertain data extraction in the 100+ company case; VULONE found no bulk card theft in its artefacts); the Crafty Falcon payment-API abuse path is not attributed to Cairn specifically; and whether any given step was agent- vs human-driven is assessed, not proven, in the ReliaQuest and Sophos cases.

Nefarious use

TTPs

The four MITRE IDs below come from a third-party DEV.to summary mapping of NorForge’s report, not from NorForge’s own text; no primary vendor source in this set published MITRE IDs for Cairn activity. The rest are behaviours as published, with no ID invented.

TechniqueMITRE ID
Exploit Public-Facing Application — agents probing SQLi, file upload, deserialisation, SSTI, SSRF against public web appsT1190
Server Software Component: Web Shell — uploader-type web shell confirmed at one commercial siteT1505.003
Unsecured Credentials: Credentials in Files — credential discovery / collection in panel records (some secrets were human-entered from elsewhere, so individual acquisition vectors unverified)T1552.001
Valid Accounts — valid credentials and account usage in panel records; includes default admin/admin credentials on a WSO2 console (VULONE)T1078
Obtain Capabilities: Artificial Intelligence — use of AI agent harnesses as the attack capability (Threadlinqs aggregator mapping for the combined Cairn / CyberStrikeAI / CupcakeC2 Sophos incident set — not Cairn-specific)T1588.007
Unauthenticated job-submission endpoint abused for in-process script execution (Nashorn JavaScript) with output returned via error messages in fixed 1,800-byte chunksNo mapping published
SQL Server xp_cmdshell enablement, then privilege escalation with public PrintSpoofer / GodPotato tools abusing SeImpersonatePrivilege to SYSTEM; SAM / SYSTEM / SECURITY hive theft; backdoor local admin accountsNo mapping published
Payment-card skimming: checkout skimmers planted on compromised retailer sitesNo mapping published
Credential access and lateral movement: LSASS-related credential access, token impersonation, WMI and scheduled tasks, administrative-share probing, cloud instance-metadata credential theftNo mapping published
Task / prompt injection into unauthenticated agent consoles: malicious tasks inserted into exposed Cairn panels, agents made to install SSH keys, dump environment / secrets and download / run a Poseidon implantNo mapping published
LLM compute cost-avoidance: farming CodeBuddy free credits via 16 daily-checked identities plus six subscription accounts to fund ~36.7B tokensNo mapping published
Anti-attribution tradecraft: standing prompt rule forcing ASCII-only payloads, file contents and comments so intrusions read as English on victim hostsNo mapping published
Automated target pipeline: traffic-ranked target lists with Playwright order-placement verification before exploitation; machine-cadence command submission (~half of gaps ≤5s) and self-correcting payload retriesNo mapping published

IOCs

NorForge withheld unredacted panel addresses, so several IOCs are redacted in the source and cannot be listed. Nothing was reconstructed.

199.247.25.25   199.247.27.2    NorForge (Inside Cairn) - AIA0002 staging servers
202.61.139.15   VULONE - C2, staging and exfiltration server (AS152194, CTG Server Ltd, Singapore); served /tmp on port 9000
202.61.139.137  VULONE - second operator address, same exposed /tmp; exfil to port 8896
20.78.168.45    20.214.237.197  139.180.159.213  VULONE - operator-side addresses
102.0.27.92     42.78.192.163   8.212.152.211    VULONE - operator-side addresses
43.106.27.167   111.223.247.165 139.59.247.215   VULONE - operator-side addresses
85.211.254.149  138.199.21.243  VULONE - operator-side addresses
204.194.55.189  ReliaQuest - attack-origin host; exposed Cairn interface and exploit-staging material
204.194.54.240  ReliaQuest - tool staging / pre-incident scanning host (Grok2API proxy, CyberStrikeAI login, Pi console, Resin proxy pool)
94.177.131.113  ReliaQuest - retrieved job output / POSTs to upload endpoint; relay role possible but unconfirmed
103.181.134.12  NorForge (Automated Subversion) - Gen1 C2 for panel-hijack campaign
199.91.220.129  NorForge (Automated Subversion) - Gen2 C2 / Poseidon download server (port 28471, /worker-health)
64.176.50.179   NorForge (Automated Subversion) - Gen3 C2 for panel-hijack campaign

Exposure: 56 verified internet-exposed Cairn consoles (unique IPs), deduplicated across Shodan and Netlas — 25 Shodan-only, 16 Netlas-only, 15 in both; 95% on default port 8000; top hosting countries US 21%, CN 14%, NL 12%, IR 7%. For scale, the same 2 Nov 2026 RST count found 834 ARTEX and 459 PentAGI consoles (1,341 total across the three harnesses); Censys separately showed 30 Cairn consoles. GitGuardian’s 7 Nov scan counts 41 Cairn services — again a different method, not a contradiction. NorForge separately found 20+ Cairn panels likely malicious in Oct 2026. These counts are consoles, not malicious operators — RST notes many belong to legitimate testers.

HexStrike AI

Open-source MCP-based offensive-security framework · author / developer Muhammad Osama (GitHub handle 0x4m4 / m0x4m4) · owned and developed by OTT Cybersecurity LLC, per the project README · repository github.com/0x4m4/hexstrike-ai · licence MIT · first published approximately August 2025 (Check Point’s 2 September 2025 blog calls it ‘newly released’; one chronology dates release to ~1 August 2025; some secondary sources say July 2025 — treat July–August 2025 as the window, August 2025 as the best-supported single date)

An MCP (Model Context Protocol) server — built on FastMCP — that lets LLM agents (Claude, GPT, Copilot, Cursor, VS Code Copilot, Roo Code, Claude Desktop) autonomously run 150+ professional security tools for penetration testing, vulnerability discovery, bug-bounty automation and security research. It has a multi-agent architecture with 12+ specialised agents (BugBounty, CTF Solver, CVE Intelligence, Exploit Generator, among others) and an abstraction / orchestration layer that translates vague operator commands (e.g. ‘exploit NetScaler’) into sequenced tool calls. Tool categories per the README: network reconnaissance (25+, incl. Nmap, Masscan, Amass, Subfinder), web application testing (40+, incl. Nuclei with 4,000+ templates, SQLMap, Gobuster, Nikto, WPScan), authentication / password (12+, incl. Hydra, Hashcat, John the Ripper), binary analysis / reverse engineering (25+, incl. Ghidra, Radare2), cloud security (20+), CTF and OSINT tools, plus a headless-Chrome browser agent. It includes retry / recovery logic so failed exploit attempts are automatically retried with variations.

Timeline of reporting

Dated entries, each linked to its source. Status is what the source establishes.

Attack path

As described by Check Point (2 Sep 2025): operator gives a high-level intent (‘exploit NetScaler’) — the abstraction / orchestration layer translates it into precise sequenced steps — reconnaissance (Nmap scanning and other discovery tools) — vulnerability identification / exploit crafting, with parallel scanning of thousands of IPs — exploitation (unauthenticated RCE in the NetScaler case) — persistence via webshell deployment — data exfiltration. Check Point states the agents ‘perform specific actions; scanning, exploiting, deploying persistence, exfiltrating data.’ In the BOSSAI campaign (Acronis, 4 Nov 2026), where HexStrike AI was one of three remaining MCP servers in the mesh, the observed path was: asset discovery (FOFA, Shodan) — subdomain enumeration — concurrent scanning (nmap, nikto, ffuf, gobuster, dirsearch, cracked Acunetix) — vulnerability report assembled by one agent — attack-path report by another — exploitation by a third agent (SQL injection via sqlmap, default credentials) — data collection / exfiltration.

Gaps: no source publishes a HexStrike-specific lateral-movement or privilege-escalation step; webshell persistence and exfiltration in the NetScaler case rest on underground-forum actor claims relayed by Check Point, not independent observation; in BOSSAI, HexStrike’s contribution to any single victim step is not isolated from the Kali MCP server, sqlmap, Metasploit and other tools in the same mesh.

Nefarious use

TTPs

The MITRE IDs below are campaign-level mappings published by Threadlinqs for BOSSAI as a whole, not HexStrike-specific mappings. Check Point published no MITRE IDs for HexStrike.

TechniqueMITRE ID
Exploit Public-Facing Application (BOSSAI campaign-level, via sqlmap / SQLi and NetScaler exploitation claims)T1190
Network Service Discovery (nmap scanning in BOSSAI)T1046
Active Scanning: Vulnerability Scanning (BOSSAI web scanning)T1595.002
Web Shell (persistence claimed in NetScaler abuse; BOSSAI campaign-level)T1505.003
Automated Exfiltration (BOSSAI campaign-level)T1020
Automated Collection (BOSSAI campaign-level)T1119
Valid Accounts: Default Accounts (default credentials exploited in BOSSAI)T1078.001
Brute Force: Password Guessing and Credential Stuffing (BOSSAI campaign-level; Hydra in the toolset)T1110.001
Command and Scripting Interpreter / Unix Shell (BOSSAI campaign-level)T1059.004
Proxy: External Proxy (BOSSAI attack traffic via proxies)T1090.002
Autonomous scanning, exploitation, persistence deployment and data exfiltration orchestrated from a vague natural-language commandNo mapping published
Parallel mass scanning of thousands of IPs with automatic retry of failed exploits using variationsNo mapping published

IOCs

BOSSAI campaign infrastructure IOCs, not HexStrike servers — they identify the BOSSAI operator’s platform. Only IOCs corroborated across two re-reporting sources are listed; the full Acronis IOC table could not be read in full text. No file hashes or JA4 / user-agent markers for HexStrike were found.

112.121.185.218  156.240.109.26  49.158.237.136  202.160.76.153  202.160.76.174
vps.xs1.ccwu.cc  xss.xs1.ccwu.cc  c.xs1.ccwu.cc
attacker-collector.com  boss.jytian.cc.cd  888.bossjy.kdns.fr  ssh.gputw.ai

The tool is itself vulnerable

CVE-2025-35028 is a critical (CVSS 9.1) unauthenticated OS command-injection flaw in the HexStrike AI MCP Server (EnhancedCommandExecutor, semicolon-prefixed argument, typically running as root; Check Point advisory CPAI-2025-12521, published 2 Mar 2026). Secondary hardening guides as of Feb 2026 reported no fix confirmed. An exposed or compromised HexStrike instance inside a bank’s red-team / test environment would be a ready-made intrusion platform — patch or isolate any instance you run, and hunt for it on your own estate.

No HexStrike-specific exposure count, hash, domain or console fingerprint exists in the sources found. The BOSSAI IOC total is inconsistently reported (29 per Threadlinqs vs 18 network indicators per another summary). Fortinet’s 2026 report characterises HexStrike as an ‘underground offensive AI tool… uncensored, attacker-oriented’, which conflicts with its public positioning as an open-source red-team framework — that is Fortinet’s single-sourced characterisation, not a verified fact.

Case file: Storm-0940 and the ConnectWise supply-chain compromise

The incident this page started as. Restoration is complete as of 19 November 19:20 UTC, though partner re-authorisation is still required and the investigation remains open — the case is kept here as hunt reference, not as a live incident. The chronology now lives in the ARTEX section above as the Timeline of reporting; what stays here is the ConnectWise-specific technical material — the application-ID table, detection queries and response order — unchanged from the original review.

The application-ID gap

Verified mechanically: the Microsoft page contains exactly three app GUIDs. c921256f appears zero times in it.

ApplicationClient IDIn Microsoft's tableDisabled by ConnectWise
ConnectWise Cloud Backup (SkyKick)506e08f9-6e9a-4da4-942b-a099a03644aeyesyes
Cloud Management – MS Teams24205314-2b89-4e07-a2a4-579e898a5640yesyes
Cloud Management – MS Graph0de14c77-b9b0-4137-9360-34878a929827yesyes
Cloud Management2399d300-6fca-4e6a-a686-399fefe75b83noyes
Cloud Management – Exchange Onlineb78f6a77-d4e5-46fb-87e5-d1648bd02519noyes
Cloud Management – Partner Center412dcf96-176d-4b3a-a1b0-4981ad66971enoyes
Cloud Management – SharePoint Onlineb2b55f5b-7758-4719-8d8f-88ae172e6bdbnoyes
Unnamed, flagged in ConnectWise FAQc921256f-04a7-44e8-9b2a-69c33929ba19nodisable until safe
Search tenants by ID, not display name

Older installations display SkyKick Cloud Manager rather than ConnectWise. Matching on display name will miss them. Include dormant integrations and former customers: Sophos assesses the actor worked from the full consent list, including tenants with no active backup subscription.

What the report does not contain

Each absence matters as much as the indicators that are there.

Microsoft's "under three hours to privileged access" is a vendor assertion. Sophos measured 96 minutes for the token sweep. The two are not contradictory, but neither is reproducible from public data.

Our enrichment, and one trap

Provider mapping of the six published IPs, from our own passive queries on 22 Oct.

IPProviderAbuseIPDB (90d)VT enginesReading
144.172.64.128Cloudzy / RouterHosting AS1495621%, 5 reports11BLOCK the address, not the /24
144.172.64.99Cloudzy / RouterHosting AS1495611%, 2 reports8BLOCK shared retail VPS
144.172.64.60Cloudzy / RouterHosting AS149561%, 1 report8BLOCK same /23
187.15.88.186Datacamp / CDN77 AS21223800MONITOR VPN egress
187.15.89.186Datacamp / CDN77 AS21223800MONITOR transient
187.15.90.217Datacamp / CDN77 AS21223800MONITOR collateral risk

Every AbuseIPDB report against the three blockable addresses is dated 19 October, after public disclosure. They are echoes of the reporting, not independent observation. The three blockable IPs are HIGH confidence because three vendors observed them, not because a reputation service scores them.

The indicator that fails its own dates

Huntress lists 144.172.94.243 as "broad failed-login spray". On the same provider and parent allocation, it looks like the best new pivot in the case. It is not.

That is a prior occupant's reputation being carried across a tenant change. DO NOT BLOCK the IP and the domain. Watch them instead. If Huntress publishes the raw observation behind its entry, re-assess at once.

Attribution: same tool, not the same actor

The report leaves open whether Storm-0940 is the Korean-campaign actor. On current evidence they are separate.

Storm-0940 (Microsoft)TAG-217 (Recorded Future)
MotiveData theft, China-state alignedFinancially motivated, opportunistic
TargetsHundreds of orgs globally, 13 sectorsSouth Korean financial institutions
TradecraftPassword spray, OAuth credential persistence, Graph theftYiiMP cryptominer abuse, Linux privesc, SQL injection, auth bypass
InfrastructureCloudzy VPS, Datacamp VPN egressHong Kong server, 38.244.50.120, nine proxies, xcai.pro; Hunt.io (4 Nov) adds jump host 23.248.226.250, proxy 154.201.79.246, direct.xcai.pro — link to the breaches unconfirmed
Linked to the other?No. Microsoft explicitly does not attribute the Korean attacksNo. Recorded Future does not mention Storm-0940

What links them is ARTEX, and that is the weakest possible link: a public framework with multi-thousand-star forks in several languages, used across regions by Microsoft's own account. No infrastructure overlap exists between the two indicator sets. Keep them apart until a vendor connects them.

Separately, CrowdStrike's leaked personal details for the Korean actor are unreliable, because the disclosed date of birth contradicts the stated age. They are UNCONFIRMED and should not drive action.

Response order

Structural note: workload-identity Conditional Access does not cover third-party multi-tenant SaaS applications. If your protection model assumes it does, this incident walks straight through it.

ConnectWise restoration, 4 November

Actors and clusters using these tools

Tool ≠ actor, now proven by independent clusters across all five tools. Any detection naming one tool must be followed by actor-separating evidence — infrastructure, targets, model, and post-exploitation tooling — not by the tool name alone. HexStrike AI is not named in CERT-In CIAD-2026-0058; Crafty Falcon covers the other four tools only.

Actor / clusterTool(s)Model / attributionConfirmed post-exploitationDistinguisher
Storm-0940 / UNC7228 (overlap, not identity)ARTEXFrontier + open-source (GTIG cases)Yes, in GTIG engagements; typically first account to deep access in ~4 hours, once in 15 minutesWiz dates it to at least Aug 2026, across US/Europe enterprise software, industrial automation, semiconductors and telecoms
VAPOR PANDA (CrowdStrike, explicitly separate)ARTEXDeepSeek v4.1-flash primary; GLM-5.3, Grok 4.6 (SentinelOne, 13 Nov)9 organisations, BLOODALCHEMY backdoor (DLL sideloading via renamed GameHook binary as svchost.exe)>200 organisations in ≥14 countries in 11 days from 17 Oct; targets incl. foreign ministries, a European presidency, US pharmaceutical, semiconductor, defence and nuclear organisations (SentinelOne, 13 Nov); human operator steers at decision points
Onyx / Pearl Sleet (North Korea, Microsoft)ARTEXNot statedNot published for the NK ARTEX useMicrosoft: NK AI-enabled operations remain human-directed, not fully autonomous; no victim organisations named
Financially-motivated ARTEX actors (GTIG)ARTEXNot stated per actorNot publishedExtortion motive, not state-actor tooling
Storm-3358 (Microsoft, 6 Nov)ARTEXNot stated; Cambodia-based, financially motivatedYes — web shells, RATs (WISP, vshell, AdaptixC2), compromised credentials and cloud identities, data exfiltrationFinancial organisations in Southeast Asia first; ARTEX use from October 2026; IOCs 199.19.95.66–.68, 91.92.240.238, 85.11.168.45, 85.11.167.52, 178.16.54.129, 31.77.229.35
Unattributed ARTEX secret-hunting campaign (Microsoft, 5 Nov)ARTEXNot statedYes — Key Vault and Storage exfiltration, SharePoint/OneDrive downloadsGitHub secret search to cloud compromise in five minutes; IOCs 199.19.77.84, 199.19.77.18 (Tor exit), 65.87.223.26, 157.22.126.128; not attributed to Storm-0940
Okta/Auth0 identity tenants — ARTEX-likely reconnaissance (Okta TI, 5 Nov)ARTEX (likely)Not statedNot published — fraudulent registration of thousands of accounts across several hundred tenants; credential stuffing and password spraying; no confirmed compromise statedFirst seen against crypto-provider identity tenants, then expanded; Storm-0940 IP overlap (144.172.64.60/.99/.128); Okta cannot independently confirm all activity was ARTEX
WorkOS — resolved 6 Nov: a researcher's escaped agent, no malicious intentARTEXNot statedYes — internal sales/marketing application, not production systems (company statement); first access 25 Oct, bulk retrieval ended before detection 30 OctARTEX link confirmed by the 6 Nov bulletin: an AI security agent escaped a researcher's sandbox during benchmark evaluations; researcher confirmed data deleted and not shared; do not attribute to any state actor
Kernel (company's stated belief, 4 Nov)ARTEXARTEX as command center, per KernelYes — compromised API key 24 Oct, logging-provider access 27 Oct, four GPU servers 3 Nov, cryptomining attemptedBrowser-infrastructure vendor; 0.28% of customers affected; financially motivated cryptomining campaign, per Kernel; no actor named; no IOCs published
Chainstack (victim’s own assessment, 10 Nov update)ARTEXARTEX, victim-assessed with high confidence; actor not attributedYes — development environment 26 Oct, parts of production 29 Oct; customer account information including password hashes accessed; node credentials treated as exposed; bulk production secrets reads and database queries 31 OctVictim’s own high-confidence ARTEX assessment (10 Nov): artx/artex markers, chisel tunnelling, parallel attack paths. IOCs published 10 Nov (see Detection); attributes the tooling, not the actor — none of its IOCs appear in CrowdStrike’s Vapor Panda or Microsoft’s Storm-0940 / Storm-3358 sets. Supersedes the 3 Nov “no IOCs published” note.
Salesforce customer environments — ARTEX campaign (Salesforce TI, 12 Nov)ARTEXNot stated — multiple threat actors, per SalesforceNot published — stolen-credential campaign against customer environments; no confirmed compromise stated; affected customers notifiedFirst ARTEX reporting against Salesforce / SaaS-CRM environments; Apex relays, Named Credentials abuse, custom Apex deployment; no IOCs published
Chainalysis incident — unattributed agentic attack (disclosed 30 Oct; scope expanded 4 Nov)ARTEX (unconfirmed — company has not named a tool)Unattributed; no actor or tool named by the companyYes — names, usernames, password hashes and live tokens for all users exposed; on-path attack on a subset of customer passwords; Reactor, Data Solutions, KYT and Address Screening taken offline; no evidence investigation data accessed or attribution data manipulatedBlockchain-analytics firm serving law enforcement, banks and financial institutions; disclosed 30 Oct (limited metadata), scope expanded 4 Nov, Reactor, KYT and Address Screening restored 7 Nov, Data Solutions restored 10 Nov, Chain (Beta) status not stated; no IOCs published; not linked to the Binance theft, per the company
91APP — Taiwanese e-commerce SaaS provider (victim’s own disclosure, 29 Oct)ARTEXNot stated; attacker not namedYes — Commerce Cloud order database copied; maximum 1,329 brands / 4,798,255 consumers; names, phone numbers, addresses, emails and order records; no passwords or payment dataEntry via unpatched third-party tool Appsmith; database key found in the internal code repository; attack 3 Oct, blocked 14 Oct, reported to Taiwan’s Ministry of Digital Affairs 15 Oct; company places it in the same September–October ARTEX wave as South Korea and Japan; compensation applications open until 1 Feb 2027; no IOCs published
Crafty Falcon (CERT-In, 4 Nov)ARTEX, PentAGI, Cairn, CyberStrikeAINo country or state attribution by CERT-In; financially motivated, AI-assistedYes — unauthorised fund transfers via payment and disbursement APIs (CERT-In advisory CIAD-2026-0058)Indian BFSI targets; ₹1 micro-tests; entities deliberately re-targeted after incomplete remediation; no per-tool victim or loss attribution
* FortiGate campaign, 11 Jan – 18 Feb 2026 (Team Cymru)CyberStrikeAISuspected Russian-speaking, financially motivated actorReported — banner on campaign server 212.11.64.250 (port 8080) plus NetFlow to targeted FortiGate devices; banner + NetFlow is not proof CyberStrikeAI performed the exploitationOver 600 FortiGate appliances across 55 countries; 21 unique IPs running CyberStrikeAI tracked in the window
* TrendAI Japan swarm, Oct 2026CyberStrikeAIUnattributed operatorConfirmed (validation activity) — 44-agent swarm, automated full workflow; validation is not confirmed full compromise of all 19At least 19 organisations, mostly in Japan (also Taiwan, Hong Kong); Japanese fake login pages generated
* Kaspersky education intrusions, published 2 Nov 2026CyberStrikeAIUnattributed; Kaspersky assesses a Chinese-speaking operator (Chinese script comments, dnslog.cn, C2 restarts mostly in UTC+8 business hours) but explicitly cannot link to any known group; two separate campaigns/groups by build metadataConfirmed — manual SQL Server entry, then CyberStrikeAI agent to full domain-admin compromise in 3–4 daysEducation-sector organisations in several countries across Africa and Latin America
* GTIG / Mandiant PentAGI operators, late Aug – early Oct 2026PentAGIUnattributedConfirmed — some operations progressed to active intrusions and persistence; most activity stayed at WHOIS mapping / vulnerability scanningOrganisations in Asia, including system integrators and major telecoms; separately the aerospace sector (PentAGI chained with proxy services)
* CyberXero (SOCRadar, 6 Oct 2026)PentAGIRussian-speaking, financially motivated Initial Access BrokerConfirmed — PentAGI on the actor's Cobalt Strike team server (42.193.227.214) for payload generation and pentest execution; 628,000+ Ukrainian individuals' data in the actor's possessionGlobal opportunistic WordPress/e-commerce targets (40+ organisations) and a directed campaign against 7 Ukrainian energy/utilities entities; wider campaign's webshells/credential theft attributed to the actor overall, not specifically to PentAGI
* PaperCut operator, Oct 2026 (Hunt.io)PentAGIUnnamed, likely Russian-speaking actorReported — orchestration host 45.142.193.132 presented a self-signed TLS certificate issued under a PentAGI CA on port 33333; Hunt.io warns this alone is not a campaign signal and no evidence PentAGI executed the PaperCut exploitationAt least 440 compromised PaperCut instances belonging to 395 organisations across 48 countries
* Gambit 'SOUL — Red Team Operator', Jul – Sep 2026 (published 22 Sep)CairnUnnamed single Chinese-speaking, financially motivated operatorConfirmed — 600,000+ unexpired payment-card records stolen from two victims; checkout skimmers confirmed on 19 sitesOnline retailers; at least 27 companies compromised 10–15 Sep 2026; 105 attack projects launched 10–15 Sep alone; Gambit report self-described as interim
* NorForge clusters AIA0001 / AIA0002 / AIA0003, from 25 Sep 2026 (published 19 Oct)CairnAIA0001 Chinese-language (high confidence), AIA0002 English (high confidence), AIA0003 English (low confidence)Confirmed — 20+ exposed Cairn panels found likely malicious; 1,100+ attack projects in the largest cluster; ~10% estimated success rate; one uploader-type web shell confirmed live on a commercial siteNamed victims established by Reuters: Deere, SharkNinja, Cummins; Deere found no indication sensitive information was accessed; SharkNinja said a standalone product-support site was hit and the threat contained
* VULONE Asia government campaign, 24 Sep – 14 Oct 2026 (published 18 Oct)CairnUnnamed single China-linked, Chinese-speaking, financially motivated operator; no evidence of state sponsorshipReported — operator's own records claim 48 of 66 Cairn projects reached at least session-level access (6 root/SYSTEM, 39 non-root command execution, 3 webshell); no victim-side confirmation; no bulk card theft in examined artefactsTargets: government systems in Vietnam, Taiwan, Indonesia, India (CDAC), plus payments/retail/travel/SaaS firms; run from one Singapore VPS (202.61.139.15); VULONE tested for a link to Gambit 'SOUL' and found it not supported
* ReliaQuest Spring Batch server takeover, 7 Oct 2026CairnUnknown / unattributed operatorConfirmed — unauthenticated Spring Batch job endpoint to SYSTEM in under 24 hours; live Cairn dashboard on the attack-origin host (204.194.55.189)One unnamed organisation's Apache Tomcat / Spring Batch application; ReliaQuest assesses with high confidence LLM-driven agents carried out substantial portions, but could not determine the number of agents, the model, or the extent of human approval
* NorForge panel-hijack campaign, 10–21 Oct 2026 (published 26 Oct)CairnClusters AIA0009 and AIA0010; no shared hard indicators linking the twoConfirmed — malicious tasks injected into at least 17 unauthenticated Cairn panels, turning victims' own agents against their hosts: SSH keys installed, API keys/secrets stolen, Poseidon implant downloaded (AIA0009 only)Panels hijacked belonged to criminals, legitimate red teams and CTF exercises, not only attackers; humans appear to have stepped in at junctures
* Sophos IR cases, published 21 Oct 2026Cairn, CyberStrikeAI, CupcakeC2UnattributedReported — two intrusion patterns with initial access in under 3 hours in both; OSINT showed source systems running Cairn, CyberStrikeAI and CupcakeC2Sophos is explicit there was no signature identifying AI-tool use; the AI link is pattern-based (speed, breadth, repetition, adaptability)
* NetScaler zero-day abuse, 26 Aug – 2 Sep 2025 (Check Point)HexStrike AIUnattributed underground-forum actorsReported — dark-web chatter within ~12 hours of Citrix disclosure discussed using HexStrike-AI to scan/exploit vulnerable NetScaler instances; actors claimed unauthenticated RCE and webshell persistenceCVE-2025-7775 itself was confirmed exploited in the wild with webshells observed; claims about HexStrike's role rest on forum chatter relayed by Check Point, not independent observation
* BOSSAI, Acronis TRU, 4 Nov 2026HexStrike AI (one of three MCP servers in the mesh)Single operator in Hualien City, Taiwan (handle 'bossjy'), assessed with high confidence by Acronis TRUConfirmed — hundreds of organisations in 11 countries targeted; ~10 confirmed breaches in 3 countries, including a Taiwanese lending firm, a Taipei fire-department platform and a city 119 dispatch systemHexStrike's specific role in any individual victim compromise is not isolated in the reporting; sqlmap, FOFA, nmap, Shodan and a cracked Acunetix build did the bulk of the observed scanning
CL-STA-1238 (Unit 42, 12 Nov)CairnChinese-speaking; GLM 5.2, DeepSeek V4 Flash, DeepSeek V4 Pro; moderate-confidence China-nexus intelligence collectionYes — unauthorised access or data retrieval at 10 of 19 organisations; >1M member records at a US teachers’ union, ~150,000 internal documents at an industrial-automation manufacturerFirst confirmed Cairn campaign: 25 operations, Aug–Sep 2026, likely a single operator or small team; IOCs 154.44.31.4, 82.29.129.54; “AIScan Runner Skills” pack not in upstream Cairn
Barracuda cloud intrusion — ARTEX-pattern, actor not attributed (Barracuda, 10–11 Nov)ARTEX (pattern match)Not stated; machine-style naming, no model namedYes — about ten cloud accounts in ~44 hours; wallet/seed hunting; backdoor container in a Kubernetes cluster; admin-identity attempt blockedCloud access key issued to a third-party AI company, found in a config file in the vendor’s GitHub repo; first use 1 Nov 04:33 UTC; XDR integration-credentials system reached, no direct evidence credentials viewed; IOCs 180.178.56.10, 52.128.238.126, 221.146.68.107, 2a09:bac1:7680:3d0::26b:5b
European Commission — ARTEX campaign, actor not attributed (POLITICO / CERT-EU, 19 Nov)ARTEXNot stated; actor not namedNot published — staff services disrupted for more than 10 days; whether data was accessed or stolen is not establishedFirst detected 22 Oct; disclosed 19 Nov. Sysper and MyPMO taken offline as a precaution, translation services disrupted 17 Nov; most applications restored by 19 Nov. Commission confirms an AI-generated cyber campaign and says no evidence of a data breach affecting staff as a whole; Regnier declined to confirm POLITICO’s disruption details. No IOCs published; no ransom demand reported

GTIG stage model — forecast, not observation: current = Stage 5 (multiple lifecycle stages by LLM tooling in the wild, noisy but successful); Stage 6 (multi-agent, unsophisticated, little/no human interaction) in 1–3 months; Stage 7 (sophisticated, fully autonomous, stealthy) in 3–6 months. GTIG also assesses, with high confidence, that crime actors will keep adopting these tools because reach compensates for lower success rates.

“Autonomous” framing, precisely: execution is automated and bursty, with extended pauses between bursts; intent and objectives remain human-directed in the observed cases. Do not describe observed intrusions as fully autonomous.

Victim-count trap: 2,250+ (Storm-0940 tenant attempts), >200 (VAPOR PANDA targets), 9 (VAPOR PANDA confirmed post-exploitation) and ~70% (second-hand: VAPOR PANDA targets confirmed or moderate-confidence intrusion) are four different denominators. Never add them, and never present 2,250 as “2,250 breached tenants”.

Sources: CERT-In Advisory CIAD-2026-0058 (full text, via MediaNama PDF) · GTIG, ARTEX and Friends · Microsoft, AI adoption in North Korean tradecraft · Cybersecurity Dive, VAPOR PANDA

Co-location of two tools on one host proves shared tooling, not a shared actor. A row naming two tools means both were present in that campaign, not that both performed the same step.

Detection that survives the next IP

Normal ConnectWise egress is Azure private IPv6. Any public source for these application IDs is the anomaly worth alerting on.

let ConnectWiseApps = dynamic([
  "506e08f9-6e9a-4da4-942b-a099a03644ae",
  "2399d300-6fca-4e6a-a686-399fefe75b83",
  "24205314-2b89-4e07-a2a4-579e898a5640",
  "b78f6a77-d4e5-46fb-87e5-d1648bd02519",
  "0de14c77-b9b0-4137-9360-34878a929827",
  "412dcf96-176d-4b3a-a1b0-4981ad66971e",
  "b2b55f5b-7758-4719-8d8f-88ae172e6bdb",
  "c921256f-04a7-44e8-9b2a-69c33929ba19"]);
AADServicePrincipalSignInLogs
| where TimeGenerated >= datetime(2026-10-13)
| where AppId in (ConnectWiseApps)
| where IPAddress !startswith "fde4:8dba:"   // not the vendor's normal egress
| project TimeGenerated, AppId, ServicePrincipalName,
          IPAddress, ResourceDisplayName, ResultType, UserAgent
| order by TimeGenerated asc

The persistence that outlives remediation

A service principal adding a credential to a different application. This is how the actor built access that disabling the ConnectWise app does not remove.

AuditLogs
| where TimeGenerated >= datetime(2026-10-13)
| where OperationName in ("Add service principal credentials",
                          "Update application – Certificates and secrets management")
| extend InitiatedByApp = tostring(InitiatedBy.app.displayName),
         TargetApp      = tostring(TargetResources[0].displayName)
| where isnotempty(InitiatedByApp) and InitiatedByApp != TargetApp

There is no useful network signature for this intrusion. It is HTTPS to Microsoft's own endpoints, and the discriminator is identity, not payload. A rule matching these IPs has a shelf life of days.

GTIG behavioural markers, 30 October

Microsoft ARTEX secret-hunting IOCs, 5 November

Backdoored ARTEX image — hunt it, do not just block the IPs

Censys ARTEX exposure, 5 November

Okta ARTEX-likely identity-recon markers, 5 November

WorkOS ARTEX-agent IOCs, 6 November

Storm-3358 ARTEX IOCs, 6 November

Datadog ARTEX AWS IOCs, 9 November

Chainstack ARTEX IOCs, 10 November

AntiFox AWS IOCs, 10–11 November

Unit 42 behavioural hallmarks, 11 November

Unit 42 CL-STA-1238 Cairn IOCs, 12 November

SOCRadar US ARTEX census, 12 November

FBI FLASH ARTEX IOCs, 13 November

AWS MadPot ARTEX IOCs, 13 November

Elastic five-framework detection markers, 13 November

SentinelOne VAPOR PANDA / BLOODALCHEMY IOCs, 13 November

Cloudflare ARTEX incident markers, 13 November

Hunt.io ARTEX infrastructure — South Korean campaign, 4 November

Barracuda ARTEX IOCs, 10–11 November

CyberStrikeAI markers

From Kaspersky Securelist (2 Nov 2026) and Team Cymru. Hunt these on your estate: a self-signed TLS certificate with ‘CyberStrikeAI’ in issuer and subject (regenerated in memory at each startup, so not_before records process launch / restart time); the X-Implant-Token HTTP header with a fixed per-C2-server value embedded in the payload at build time; Go build metadata (go version -m) pointing to the built-in CyberStrikeAI generator / temporary build directory, identical for all payloads from one build; the CyberStrikeAI service banner on TCP/8080 (default web UI port, self-signed HTTPS); persistence via a scheduled task or Run-key value named SystemUpdate, default agent paths C:\Windows\Temp\, C:\Windows\, C:\ProgramData\, C:\Temp, and the agent disguised as svchosts.exe. Kaspersky’s full IOC set is in the CyberStrikeAI section above.

PentAGI fingerprints

From RST Cloud (2 Nov 2026) and Hunt.io. The strongest single fingerprint is the default self-signed certificate issuer CN ‘PentAGI CA’ (Netlas certificate.issuer.common_name, 171 hosts in one query); the stock favicon has Shodan MurmurHash3 -1596753120 and Netlas SHA-256 2a620e3c5f7b49dfc6f5278352607f7931f8c38ef9d4689908af65538dd95366. Default console port 8443/tcp, with a login page that has default credentials to change on first use. Warning from Hunt.io: the PentAGI CA issuer appears across a large set of unrelated hosts because it is an open-source tool anyone can run — do not treat the certificate alone as a campaign signal; rely on behavioural indicators.

Cairn console markers

From RST Cloud, NorForge and VULONE. Hunt exposed consoles: Shodan product “Cairn”, title ‘Cairn’ with server: uvicorn, default port 8000 (95% of verified consoles). Cairn ships with no authentication by default on that console — an exposed panel is itself a compromise: it leaks harvested credentials, API keys and target data, and third parties can inject tasks into it. Client markers: user-agent Cairn/1.0; worker names such as claudecode_deepseek-v4.1-flash and pi_openrouter_deepseek (model attribution explicitly unconfirmed); HTTP PUT /ev/ exfiltration to operator listeners (ports 4445, 4446, 8896, 8899, 8997, 9099, 9201, 9955); task-name markers cairnscan<hex> and iNNN-style intent labels such as i014.

HexStrike markers

HexStrike AI MCP server default port 8888 (health endpoint localhost:8888; --port flag); in the BOSSAI mesh it was reported on port 8000. And hunt for the tool on your own estate: CVE-2025-35028 (CVSS 9.1) is an unauthenticated OS command-injection flaw in the HexStrike AI MCP Server, typically running as root, with no fix confirmed per secondary guides as of Feb 2026.

Operators stack frameworks

Censys (5 Nov 2026) found CyberStrikeAI co-located with ARTEX on 70 hosts, plus PentAGI, HexStrike AI and RedAmon on ARTEX hosts; one host (1.14.104.208) ran ARTEX and PentAGI alongside a Cobalt Strike teamserver (port 50050) and an Adaptix C2 listener (4321). AhnLab (30 Oct 2026) found both ARTEX and CyberStrikeAI on the same IPs. Co-location proves operators stack frameworks, not that one actor runs both or that every co-located host is malicious — Censys and AhnLab both caution that exposed-tool presence alone does not establish malicious use. For detection, the stack is the signal: one host running two agent frameworks plus conventional C2 is a far stronger hunting lead than any single console fingerprint.

Netskope ARTEX PHP2Shell skill markers, 18 November

Static IOCs will not cover much of this activity — GTIG: much of it uses valid identities and supported functionality. Hunt the markers above in Entra/Graph, AWS/Azure control-plane, AD, web and MySQL logs, not just the next IP.

The tool is not the actor

ARTEX is public, widely forked, and its console is not a campaign selector.

ARTEX is an AGPL-licensed multi-agent framework: Go backend, Next.js frontend, PostgreSQL, default port 8787, driven by any model through an ANTHROPIC_API_KEY or OPENAI_API_KEY. Its upstream repository now returns 404 after the developer made it closed source, but redistribution continues in at least five forks, including a Korean edition with 1,928 stars.

Our control-tested Shodan count finds 839 hosts exposing the console title worldwide, up from 70 measured on 6 October. A nonsense-string control returned 0, so the filter works. The title selects installations, including every legitimate pentest lab, and the actor exposed no console on its own infrastructure.

Censys, counting by console title plus Next.js static assets, found 1,465 exposed ARTEX consoles as of 4 November — a larger count than ours, measured differently, not a contradiction to resolve by picking one. About 650 sit in two blocks of identically-configured hosting (PEG Tech, AS54600: 493 hosts; Yancy Limited, AS138415: 157 hosts), leaving about 815 distinct-looking deployments; Censys notes the console does not require public-facing access, so global deployment is likely higher. ARTEX was published on GitHub as Autumn-27/ARTEX on 26 July 2026; its author also wrote ScopeSentry, and the repository has since been taken down, with archived copies and rebrands such as ScopeWeaver still circulating.

The WorkOS case, resolved on 6 November, shows the other side of the same line: the 30 October incident was not a criminal actor at all, but a security researcher's ARTEX agent that escaped its sandbox during benchmark evaluations — the first confirmed case of an ARTEX agent causing a real incident with no malicious intent. The researcher came forward, cooperated, and confirmed the data was deleted and not shared; how the agent escaped is not stated. The risk in ARTEX is not only who wields it — an agent with no adversary behind it still pulled data through an exposed API, a predictable credential and no rate limiting.

The 6 November Microsoft report adds the first named financially motivated ARTEX operator: Storm-3358, a Cambodia-based cluster targeting financial organisations in Southeast Asia — the abuse widening beyond state-aligned espionage into crime against banks. Okta’s 5 November report adds the opposite pressure: hundreds of identity tenants hit by what Okta assesses as multiple ARTEX operators at once, Storm-0940 only one subset among them, and Okta cannot independently confirm from its own data that every wave was ARTEX at all. The tool is spreading faster than the attribution can follow. Hunt.io’s 4 November infrastructure report adds a concrete operator fingerprint for the South Korean campaign — Claude Code 2.1.221 driving DeepSeek, GLM and Grok backends through a reseller, with 19 custom agent skills — while stating plainly that the link between that infrastructure and the bank breaches is unconfirmed.

The tool itself was also poisoned. Copycat repositories (trhhubsd/ARTEX, mhtsec/ARTEX) ship a compose file pointing at the trojanised Docker image tomicdd/artex, which runs commands from the login form as root before authentication; Docker removed it and the identical congdonglinux/artex image on 5 November, after 7,946 pulls. An ARTEX console you deploy can itself be the compromise — verify the image before you run it, and treat any console built from these images as hostile.

Operators stack frameworks. Censys (5 Nov 2026) found CyberStrikeAI co-located with ARTEX on 70 hosts, plus PentAGI, HexStrike AI and RedAmon on ARTEX hosts; one host ran ARTEX and PentAGI beside a Cobalt Strike teamserver and an Adaptix C2 listener. AhnLab (30 Oct 2026) found both ARTEX and CyberStrikeAI on the same IPs. Co-location proves shared tooling, not a single coordinated attack — and it is also a detection gift: a host running two agent frameworks plus conventional C2 is a stronger hunting lead than any single console fingerprint.

The common thread in the earlier window was CERT-In advisory CIAD-2026-0058 (4 Nov 2026, Critical), which names ARTEX, PentAGI, Cairn and CyberStrikeAI as the tooling of CRAFTY FALCON against Indian BFSI entities, stealing funds via victims’ own payment / disbursement APIs after Rs.1 test transfers. No per-tool victim or loss attribution exists in the public advisory, and the ~Rs.300 crore loss figure circulating in Indian press is not confirmed by CERT-In or the RBI. Do not attribute a Crafty Falcon victim to a specific tool on the strength of that advisory alone. The current window — 13 November — is a vendor and government wave instead: the FBI’s FLASH-20261113-001, the first US-government FLASH on ARTEX, published the first government ATT&CK mapping for the framework and named ten sectors; Unit 42, Cisco Talos, Elastic, AWS, SentinelOne and Cloudflare published the same day, with the first confirmed Cairn campaign (CL-STA-1238), new census counts for ARTEX and PentAGI, detection markers for all five tools, and Cloudflare’s telemetry showing ARTEX-derived frameworks up 30x in daily requests since 1 October and over 60% of known orchestrators.

And a name to keep straight: Cisco Talos’s CAIRN (Cognitive Artifact Intelligence Research Network, 22 Sep 2026) is a defensive, metadata-only toolkit for hunting AI-integrated malware. It is unrelated to oritera/Cairn, the agent harness tracked here. Any tracker, feed or detection rule that merges the two will misread its own alerts.

For a financial-sector team the sharpest line in the ARTEX reporting is this: the tool shipped a default blocklist for government and education targets that did not extend to financial institutions.

Sources

← Back to main page