ARTEX, the AI pentest agent attackers adopted: what to fix first
The FBI, Microsoft, CrowdStrike and Check Point reports on ARTEX, read in full: how often it worked, why credentials were the way in, and a log check.
ARTEX is the first AI penetration-testing agent with a long public record of being used against real organisations. In under sixteen weeks it went from a GitHub release to an FBI alert, with Microsoft, CrowdStrike, Check Point and Unit 42 publishing case work in between. We read those primary reports in full for this article, and they tell a less cinematic story than the headlines. The agent did not find new classes of bug. It found old passwords, old keys, development credentials left in production and API descriptions left in public, and it found them fast. Measured by its own campaigns, it also failed most of the time.
This post covers what ARTEX is, a dated timeline, a success-rate table built from the reported figures, where the credentials came in, what to fix first, how to look for agent traffic in your own logs (with a script), and what to tell a colleague who wants to try it.
On names: this is ARTEX the penetration-testing framework, not the insurer or anything else that shares the word.
What ARTEX is
The FBI's FLASH-20261113-001 describes it as "an open-source, large language model (LLM)-driven agentic platform" written in Go. An operator states a goal in plain language; a planner agent breaks it into tasks and worker agents run ordinary security tools to carry them out. ARTEX ships no model of its own. It calls whichever model the operator connects, and the FBI lists Anthropic and OpenAI models alongside DeepSeek, GLM, Grok and Kimi.
A developer using the handle Autumn-27 published it in July 2026; the oldest commit in the preserved history is dated 26 July. The FBI estimates it reached 1,500 to 2,000 GitHub stars in ten weeks. On 8 October, the day after CrowdStrike tied it to attacks on South Korean financial firms, the developer took the repository down, saying it had been meant for authorised testing. That did not stop it. The FBI notes that it lives on through community copies, archives and a commercial edition called ARTEX PRO, and counts more than 1,600 consoles exposed to the internet as of early November.
From release to FBI alert
| Date (2026) | Event | Reported by |
|---|---|---|
| 26 July | Oldest commit in the ARTEX history | GitHub commit record |
| Late Sept to early Oct | Campaign against South Korean financial organisations | CrowdStrike, 7 Oct |
| 8 Oct | Original repository removed by its developer | FBI |
| 10 to 14 Oct | Storm-0940 password spray: 151,000+ failed logons, no successes | Microsoft, 21 Oct |
| 17 Oct onward | VAPOR PANDA targets 200+ organisations in at least 14 countries | CrowdStrike, 29 Oct |
| 25 to 30 Oct | A researcher's ARTEX run reaches an internal WorkOS tool | WorkOS bulletin |
| 14 to 29 Oct | iOS signing-certificate theft campaign | Check Point, 5 Nov |
| 5 Nov | Docker removes two backdoored ARTEX images | StepSecurity, 12 Nov |
| 11 Nov | Unit 42 publishes incident-response findings | Unit 42 |
| 13 Nov | FBI issues FLASH-20261113-001 | FBI |
Microsoft attributes the ConnectWise Cloud Backup supply-chain compromise to Storm-0940, a China-aligned actor that used ARTEX to attempt sign-ins to more than 2,250 downstream tenants. CrowdStrike describes VAPOR PANDA as a China-nexus espionage group running several engagements at once, with a single human steering the agents at decision points in Chinese. The FBI's own summary says some intrusions went from first targeting to privileged access in under three hours.
How often it worked
The reports give enough figures to count attempts against results. The table is the dataset that comes with this article, run through its script; every row names the report it came from.
| Campaign | Attempted | Succeeded | Rate | Source |
|---|---|---|---|---|
| Storm-0940 password spray | 151,000+ failed logons | 0 sign-ins | 0% | Microsoft |
| VAPOR PANDA targeting | 200+ organisations | 9 with confirmed post-exploitation | 4.5% at most | CrowdStrike |
| iOS signing certificates | 300+ certificates stolen | 3 passed validation | 1.0% at most | Check Point |
| Gaming accounts | 190,000+ credential sets | about 6,000 valid | 3.2% at most | Check Point |
| Crypto draining | 230 agent "findings" | 0 moved funds | 0% | Check Point |
Check Point itself rounds the gaming figure to about 3.3 per cent; 6,000 of 190,000 is 3.16. Where a report says "more than", the true denominator is larger, so the rate shown is a ceiling.
A worked example: the spray that got nothing. Microsoft counts more than 151,000 failed logons against 327 accounts at eight organisations between 10 and 14 October. That is about 462 attempts per account. On 11 October alone there were more than 111,000 failures, peaking at 2,900 a minute, which is 48 every second. Forty-eight accounts took more than 1,000 failures each. Lockout responses made up nearly 98 per cent of the activity, and Microsoft saw no successful authentication, token issuance, Graph call or directory change. The report does not break down which other controls held, so the honest reading is narrow: a spray at machine speed ran into ordinary lockout and produced nothing. Our explainer on credential stuffing and rate limiting covers how those controls are tuned.
The limit of this table matters as much as its numbers. Low rates across very large attempt counts still produce victims: nine organisations breached by one group, about 6,000 live gaming accounts, and, in the FBI's count, personal data of roughly 67,000 people taken from South Korean financial institutions. Check Point's crypto campaign also shows a failure mode defenders should not count on. The agents reported 230 successes, and some of the claimed cash-outs were hallucinated, because success was judged from API responses alone. That cost the attacker time; it cost the targets nothing. Campaign costs in that report ran from about $145 to $1,500 each, which is why cheap failure at scale is still a business.
Where it got in: credentials, not new bugs
The FBI's table of observed initial access reads like a list of old audit findings: hardcoded administrator passwords in single-page-app JavaScript bundles, weak administrative credentials, missing authentication on web APIs, stolen VPN configurations and leaked credentials, account takeover through password resets, alongside SQL injection, SSRF and unsigned SAML assertions. The case reports add detail.
- Keys nobody rotated. CrowdStrike describes a Greek government target where the operator supplied a 2022 directory export holding authentication keys that were likely never rotated. The agent used them to forge an administrator token and read mailboxes, then used VPN configurations from sent mail and a VPN client key from a May 2024 email leak to get onto the network.
- Password reset as the front door. Unit 42 reports attackers reaching cloud access through public password-reset pages and exposed APIs "without exploiting a single software vulnerability", sometimes in less than a day. In one case more than 200 client secrets were validated in about three minutes, and the operator could act on a stolen credential within a minute of finding it.
- A development credential in production. WorkOS says a security researcher's agent, running ARTEX benchmark evaluations, left its test environment and reached an internal account-lookup tool. The API was reachable from the internet, accepted a weak credential intended for local development, and had no rate limit, so records came out in bulk.
- Credentials in plain files. Check Point's gaming campaign read more than 190,000 credential sets from plaintext files left in web roots.
- Secrets on GitHub. StepSecurity summarises Microsoft research from 5 November: agents ran GitHub searches pairing a target's name with words like "secret", then used the application secrets they found. In one case the path from search to Azure Key Vault data took five minutes.
None of these needs a model to discover. What the agent changes is the cost of checking every one of them, on every peripheral system, at once. Our post on stealer logs in the breach record covers the other common source of working credentials.
What to fix first
The recommendations below come from the FBI's list, with the matching points from CrowdStrike, Unit 42 and WorkOS. They are in rough order of how directly they answer the cases above.
- Treat any exposed secret as stolen and rotate it, including keys sitting in old backups and directory exports (CrowdStrike), and scan repositories, container images, build artifacts and pipelines continuously for secrets (FBI). Deleting the commit does not undo the exposure.
- Keep development credentials out of production, and add a control that blocks them at deploy time, the step WorkOS says it is now adding.
- Rate-limit and bot-protect password reset, registration and verification-code endpoints, and remove unauthenticated checks that reveal whether an account exists (Unit 42). Keep lockout and spray detection on sign-in; the Microsoft numbers show what they are worth.
- Take down the attack map. Remove public OpenAPI specifications, source maps and config files that carry tokens or client IDs, and require authentication on every write endpoint, including development and staging (Unit 42).
- Stop teaching the agent. Do not return stack traces, SQL errors, internal paths or framework details to unauthenticated users (FBI). Agents use error detail to correct their next request.
- Inventory the edges. The FBI says ARTEX does best against partner portals, mobile backends, employee and sales-support apps, and legacy services: the systems that get less attention than the main product.
- Put limits on high-impact APIs (rate, recipient and transaction thresholds) and alert on sequential lookups or unusually many records per session (FBI).
Where a session is already stolen, cookie flags and rotation decide how long it stays useful; that is covered in session tokens, cookie flags and rotation.
Finding agent traffic in your logs
The FBI's behavioural indicators are high-frequency automated requests with sequential enumeration against login and search endpoints, high volumes of failed attempts with small changes between them, and unexpected outbound calls from servers to model APIs such as api.anthropic.com, api.deepseek.com, open.bigmodel.cn and api.x.ai. That last one is the cheapest alert on the list: an application server that starts talking to a model API it has never used deserves a look.
Two vendors add sharper rules, which should be read as vendor guidance. Abstract Security describes error recovery (several 400s on a path, then a quick 200 with corrected parameters), user-agent rotation over a single TLS fingerprint, and fetching an API specification before sweeping internal routes. Kodem Security describes a model "think time" cadence of short bursts about a second apart, then a pause of 25 to 31 seconds.
The script that comes with this article turns those into six checks over an ordinary access log (Apache or nginx combined format, or IIS W3C). The core of the error-recovery rule is short:
errs = defaultdict(list)
for ts, path, status, _ in rows:
if 400 <= status < 500:
errs[path].append(ts)
elif 200 <= status < 300 and len(errs[path]) >= 3 and (ts - errs[path][-1]).total_seconds() <= 60:
recov += 1
errs[path].clear()
Run it with --log on your own file, or with --demo to see it on a small synthetic log with one ordinary visitor and one agent-like client:
python artex-ai-pentest-agent-what-defenders-should-know.py --demo
203.0.113.9 requests=93 behavioural signals=5
burst-pause 6 bursts followed by a 20-35 s pause
error-recovery 1 path(s) fixed after 3+ client errors
ua-rotation 4 user agents within 30 minutes
auth-pressure 40 login/reset/token requests in 10 min, 40 failed
map-then-sweep fetched /swagger.json, then 50 distinct paths in 10 min
The demo traffic is invented to show the output; it is not a capture of ARTEX. The limits are real ones. Access logs record time to the second and carry no TLS fingerprint, so the rotation rule is weaker here than in the vendor version. User-agent strings are set by the attacker; the script reports scanner names only as a supporting signal. Every threshold is a starting point, and authorised testers trip the same rules, so a flagged address is a lead for a person, not a block list. For where automated triage of alerts like these works and where it does not, see LLMs in the SOC.
If someone on your team wants to run it
There is no official source to download from. The original repository is gone, and what remains is a spread of copies of uneven provenance. That has already hurt the people running it. StepSecurity reports that two Docker Hub images, tomicdd/artex and congdonglinux/artex, carried a backdoor that ran commands sent through the console's login form as root before the password check. The source code matched the original, so a code review would not have caught it. The image was pulled about 9,131 times before Docker removed both on 5 November, and 637 GitHub repositories still pointed their Compose files at it. Anyone who ran either image should rebuild the host and rotate every model-provider key and every credential the console held.
The most visible English copy had its own problem. Advisory GHSA-q4j5-2wwq-gmq5, rated high and published on 5 November, covers versions before 0.5.1: an uninitialised instance could be taken over by anyone who reached it, the console was publicly exposed, and its login had no rate limit. With more than 1,600 consoles on the internet by the FBI's count, that is not a theoretical risk.
The safety controls do not settle the question either. The header of the tool's own guard/guard.go says the authorisation-scope mechanism was removed, and that the built-in intercept rules are ordinary rules users can disable or delete. Those controls can stop a careful operator's mistakes; they cannot stop an operator who switches them off. The WorkOS case shows a third risk, an agent reaching a system nobody meant it to touch. Authorisation to test comes from a written agreement with the system's owner, and the agent's traffic still counts as the operator's.
There is also the model. Every target response passes through whichever model provider, or reseller, the operator connects. CrowdStrike found the Korean campaign reaching DeepSeek through a reseller. Check Point reports that guardrails were hit only on Anthropic's Opus 5.5 among the American models configured, after which the operator switched to more permissive Chinese models, and the FBI says DeepSeek Flash models dominate attacker instances partly because of a "lack of observable guardrails" under a pentest framing. A client's data sent through an unvetted reseller is a data-handling decision in its own right.
What the record does not show
No published benchmark measures ARTEX itself, and we found no independent audit of any surviving copy. Documented legitimate use is thin: WorkOS now tests its own internet-facing systems with ARTEX, after being on the receiving end of it. Figures disagree between sources, most visibly the count of exposed consoles, which runs from 609 hosts in Unit 42's October count to 1,608 live panels in Check Point's report, depending on date and method. This article prints the figure each source gives and names the source.
What the record does show is consistent across every report we read. The agents are fast, loud and heavily dependent on the person steering them, and they win where a password, key or token was already exposed. The fixes are the ones on any hygiene list. The difference now is how quickly an untended one gets found. If one item from this post goes on next week's list, make it the first: find the secrets you have already exposed, and rotate them.
Sources
- FBI, FLASH-20261113-001, "Threat Overview of ARTEX Autonomous Penetration Testing Framework", 13 November 2026 (the summary, the threat section, the TTP table, the network and behavioural indicators and the recommendations; read in full 2026-11-18)
- Microsoft Security Blog, "AI-driven, large-scale nation-state campaign using ARTEX enables supply chain compromise", 21 October 2026 (the password spray figures, the 2,250 tenants, the under-three-hours cases; read 2026-11-18)
- CrowdStrike, "VAPOR PANDA Uses AI-Driven ARTEX to Scale a Global Espionage Campaign", 29 October 2026 (200+ organisations, nine with confirmed post-exploitation, the unrotated keys and the leaked VPN key, the recommendations; read 2026-11-18)
- CrowdStrike, "Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance", 7 October 2026 (the first public attribution, the models and the reseller; read 2026-11-18)
- Check Point Research, "Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns", 5 November 2026 (the three campaigns, their costs, the model refusal and the exposed panels; read 2026-11-18)
- Palo Alto Networks Unit 42, "ARTEX and AI-Powered Red Teaming Tools: What Defenders Need to Know", 11 November 2026 (password-reset pages and exposed APIs, client secrets validated in minutes, host counts, the recommendations; read 2026-11-18)
- WorkOS, security bulletin "Update for October 30th, 2026 data security incident" (the development credential in production, the researcher's agent, the changes WorkOS made; read 2026-11-18)
- StepSecurity, ARTEX and leaked secrets, 12 November 2026 (the backdoored Docker images, 9,131 pulls, the 637 repositories, the user agents, the advice for affected operators, and its summary of Microsoft's 5 November research on GitHub secrets; read 2026-11-18)
- GitHub security advisory GHSA-q4j5-2wwq-gmq5, "Unauthenticated admin takeover on uninitialized instances, public exposure, and missing login rate limit", high, 5 November 2026, versions before 0.5.1 (read through the GitHub API 2026-11-18)
- ARTEX source, guard/guard.go header comment on the removed authorisation-scope mechanism and the built-in intercept rules users can disable or delete (read 2026-11-18)
- GitHub API, commit history of the English copy: the oldest commit, by Autumn-27, dated 2026-07-26T08:27:45Z (read 2026-11-18)
- Kodem Security, ARTEX response runbook, 17 November 2026 (vendor guidance: the request cadence of bursts and 25 to 31 second pauses)
- Abstract Security, detection guidance for AI pentest frameworks, 10 November 2026 (vendor guidance: error recovery, user-agent rotation, specification fetch then route sweep)