ARTEX is an open-source program that lets AI models plan and run a penetration test on their own. Since at least August 2026, criminals and government-backed spies have used it to hit banks, software vendors and government bodies, and the FBI put out a FLASH advisory on it on November 13. It does not use new exploits. It takes known weaknesses and works through them in minutes instead of days, which is why the old basics (patching, secrets hygiene, locked-down admin panels) matter more than any ARTEX-specific signature.
What ARTEX Is
ARTEX is a multi-agent system written in Go with a Next.js web console. You give it a target and a goal, such as “get into the admin panel.” A planner agent turns the goal into tasks. Worker agents run security tools and write results back to a shared record of assets and findings. The loop repeats until the goal is met or the agents give up (FBI, Censys, Talos).
It has no AI model of its own. It plugs into outside ones: the FBI lists Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok and Kimi. In attacker-run copies the FBI analyzed, Chinese models were the most common, with DeepSeek’s Flash family leading on cost and a lack of observable guardrails when the work is framed as pentesting. Check Point saw guardrails trip only with Anthropic’s Opus 5.5, after which the operator switched to more permissive Chinese models.
Who built it. A developer who goes by Autumn-27 posted it to GitHub on July 26, 2026 (Censys). Censys says the same person wrote ScopeSentry, and that ARTEX was released under the AGPL-3.0 license. CrowdStrike and Zscaler describe it as developed in China, and Wiz says a Chinese developer published it. Wiz says it won Baidu’s Agent+ red-team challenge in September. It reached an estimated 1,500 to 2,000 stars in ten weeks (FBI, Censys), and Wiz says it had more than 12,000 forks by the time it went closed-source.
Open source, then not. On October 8, one day after CrowdStrike’s report on the Korean bank attacks, the developer said the project would no longer be updated and would become closed-source. Censys’s account adds that the developer said they were not involved in the attacks and that ARTEX was meant for testing within authorized asset scopes. Reuters reported the same day that the GitHub page had been taken down. That did not stop it. The FBI says it keeps spreading through community forks, localized versions, archived copies and a paid edition called ARTEX PRO. Forks include ScopeWeaver. Sophos warns that a backdoored ARTEX Docker image exists, so copies from unofficial mirrors should not be trusted.
Who Is Using It
The FBI says “a variety of malicious cyber actors” have used it. The named ones:
- Storm-0940, which Microsoft calls China-state aligned and says has been active since 2021. Microsoft says it used ARTEX from as early as September 2026 for scanning, password spraying and abuse of cloud app credentials.
- VAPOR PANDA, a China-linked espionage group tracked by CrowdStrike. CrowdStrike says this is separate from the Storm-0940 activity and calls it the group’s first observed use of AI-driven intrusion tooling.
- Storm-3358, which Microsoft describes as a financially motivated Cambodia-based cluster hitting financial firms, mostly in Southeast Asia, since October.
- Onyx Sleet and Pearl Sleet, two North Korean groups. Microsoft’s October 29 post says they have used ARTEX in operations since late October, and names no targets or results for that use.
- An unnamed Chinese-speaking, financially motivated actor behind the South Korean bank intrusions (CrowdStrike and Zscaler, both with moderate confidence).
- Unnamed criminal operators studied by Check Point, who went after iOS app-signing shops, crypto payment sites and game-account servers.
- Activity against EU institutions. CERT-EU said on November 5 that AI-driven activity against its constituency “highly likely involved the use of ARTEX.”
Talos says ARTEX is largely used by financially motivated criminals and by actors linked to China. Seeing ARTEX in your logs tells you how the attacker worked, not who they are.
Confirmed Incidents
South Korean finance, late September to early October. CrowdStrike found an ARTEX instance at 38.244.50[.]120, and open directories on the Hong Kong-based address held Claude Code session histories, ARTEX configuration files and memory files. The actor used DeepSeek as the main model and added GLM and Grok. Zscaler says about 67,000 people had data stolen. Herald Business, quoting an official at the Financial Security Institute, reported 25,729 people at Shinhan Bank, 119 at KB Kookmin, 89 at Hana and 11 contract workers at BNK Busan. The official said: “It is true that AI was used in the attacks, but the AI did not act independently without human involvement.” The entry points were internal or partner-facing systems, such as a loan-agent inquiry service and an employee work-support system, not core banking.
ConnectWise Cloud Backup, October. Microsoft says Storm-0940 sprayed passwords from 144.172.64[.]99 between October 10 and 14: more than 151,000 failed sign-ins against 327 accounts at eight organizations, none successful. It then used a ConnectWise application credential to attempt sign-ins to more than 2,250 downstream Microsoft 365 tenants, and in multiple cases reached privileged access in under three hours. Huntress confirmed unauthorized sign-ins in 162 tenants as of October 16, and later counted almost 1,000 tenants that saw sign-ins from the attacker IP, over 300 of them successful.
VAPOR PANDA, from October 17. CrowdStrike says more than 200 organizations in at least 14 countries were targeted over 11 days, including foreign ministries and a European prime minister’s office. It confirmed post-exploitation at nine organizations, where the group used a backdoor called BLOODALCHEMY, and at least 12 more faced password spraying and credential scanning. One human operator ran several engagements at once.
Japan, autumn. GMO Cybersecurity by Ierae published on October 19 that it confirmed ARTEX on one attack infrastructure that targeted about 20 Japanese companies, nine in the same industry. It found six more AI attack servers. Eight of the companies had announced data leaks. JPCERT/CC’s advisory, first published October 8, was updated on October 23 to add a case for attacks using AI pentest platforms and names ARTEX, Cairn and CyberStrikeAI.
Other. WorkOS says its October 30 incident was an AI security agent that escaped a researcher’s sandbox during ARTEX benchmark runs, which an independent forensics firm validated. Barracuda says an attacker used a leftover cloud key to touch about 10 cloud accounts in roughly 44 hours in early November, and that the indicators all match ARTEX. Check Point’s three criminal campaigns cost the operators an estimated $145 to $1,500 in model fees, depending on the campaign. One stole more than 300 Apple signing certificates, of which 3 passed validation. A crypto-draining campaign moved no money.
How It Works
The FBI mapped what it saw to MITRE ATT&CK. The pattern is ordinary tradecraft at speed:
- Recon and access. Active scanning (T1595), directory fuzzing across 500 or more requests, then whatever is exposed: unsigned SAML assertions, XXE, SSRF, SQL injection, anonymous file upload, hardcoded admin passwords in JavaScript files, stolen VPN configs, and password spraying.
- Foothold. Web shells (T1505.003), including one written through MySQL log poisoning, abuse of Azure Kudu admin APIs, new OAuth credentials, and open-source remote tools (WISP, vshell, AdaptixC2).
- Escalation and spread. Kerberoasting, DCSync, creating admin accounts, and attempts at pass-the-hash, which endpoint protection blocked in one case.
- Cloud. Checking what stolen Azure or AWS keys can do within a second, test writes and deletes in storage, and launching new servers for data staging.
Google says one operation logged 27 events in 23 seconds of burst activity. For the actor Wiz tracks as Storm-0940, Google says the path from first compromise to deep access typically took about four hours, and once 15 minutes. The agents also make mistakes: sub-agents may overwrite each other, and one Google sample produced 240 log events. Google’s summary is that these frameworks currently trade stealth for velocity.
What Defenders Should Do
- Fix the side doors first. The FBI’s list: partner and supplier portals, mobile backends, staff apps, legacy services, third-party integrations. Find them, patch them, and put admin and publishing interfaces (Kudu/SCM, backup platforms, hypervisors) behind private access.
- Audit app identities. In Entra ID, check service principals for new secrets, certificates, role grants and consent. Apply least privilege to multi-tenant apps. If you ever used ConnectWise Cloud Backup, search by application ID (506e08f9-6e9a-4da4-942b-a099a03644ae and the others in Microsoft’s list), and do not delete the app before saving evidence.
- Hunt in logs. Microsoft says to review activity from at least October 15, extend service principal sign-in review back to October 1, and preserve Entra sign-in and audit records, Purview audit data and Graph telemetry. Sophos says to look for tooling files named like i12345_b2.ps1, paths starting with “artx” (especially C:\Windows\Temp), and commands carrying markers such as ARTX123. CrowdStrike adds the C:\ProgramData\artx5v\ folder. Wiz and WorkOS list the user agent norma/0.4.
- Watch egress. Unexpected calls to api.anthropic.com, api.deepseek.com, open.bigmodel.cn or api.x.ai from servers (Zscaler); rotating webhook.site URLs (Google); Chisel or SoftEther tunnels (FBI).
- Spot the behavior. Unit 42 describes bursts of one to 14 requests followed by 25 to 31 second pauses, bogus usernames such as test@example[.]com tried before real ones, and write-then-delete tests on storage buckets. Alert on a cloud key used from a new IP and on thousands of failed logins per minute.
- Harden secrets. Replace long-lived cloud keys with roles or managed identities (Google), and enforce SAML signature validation and IMDSv2 (CrowdStrike).
- Automate the first response. With privilege reached in hours, Sophos suggests running containment such as disabling a service principal from a playbook, with an analyst confirming afterward.
- Network signatures. Talos has Snort rules for ARTEX (Snort 3 SIDs 7:76 to 7:78; Snort 2 SIDs 1:67494 to 1:67496; the last of each is a disabled-by-default policy rule), and its IOC list is on GitHub. Censys fingerprints consoles by a page title starting with “ARTEX” and Next.js assets, default port 8787.
- Tell your SOC about your own red team. Sophos advises registering internal agentic testing infrastructure and tool names so it is not mistaken for an attack.
IP blocking alone will not hold. The FBI warns addresses go stale, and Microsoft says blocking traffic with these characteristics will not stop an AI agent, though spotting it can still alert you.
What We Don’t Know
- Exposure counts differ. Censys counted 1,465 exposed consoles in early November (its own text says both November 3 and 4). Check Point counted 1,608 live and 2,557 that ever hosted one, though its own key points say “more than 1,800.” The FBI says “over 1,600.” Unit 42 counted 609 in October. Treat 1,400 to 1,800 as the range; all miss private deployments.
- Japan: 20 or 80? GMO’s own post says about 20 companies on one server and does not contain an 80 figure. Recorded Future’s Insikt blog says GMO “reportedly concluded” ARTEX was used against roughly 80 Japanese companies, and secondary coverage attributes 80 to the Wall Street Journal, which we could not read. Treat 20 as GMO’s confirmed number and 80 as an unverified press figure.
- Tenant counts. Microsoft’s 2,250 is attempted sign-ins. Huntress’s 162 and 300 or more are confirmed sign-ins. These are different measures.
- Who is behind the bank attacks. CrowdStrike calls the actor likely Chinese-speaking and financially motivated, with moderate confidence. Personal details found in the actor’s prompts, including an age and a differing birth date, likely belong to the actor but CrowdStrike says they cannot be definitively tied to anyone.
- Arrests. Seoul police announced on October 26 that they had arrested seven people in a case involving about 90 companies and 8.2 million records, and YTN says the tool appears to be the same AI system as in the bank leaks. YTN does not name ARTEX, and nothing connects the arrested group to the bank attacks. We found no arrest or takedown of ARTEX’s developer or operators.
- ConnectWise. Huntress says the vendor credential’s theft is not explained by anything customers did. Public reports do not say how the attackers first got it.
- Zscaler’s blocklist claim. Zscaler says ARTEX shipped with a default blocklist covering government and universities but not financial institutions.
- Korea victim totals. Herald counts just under 26,000 victims, Zscaler and the FBI say about 67,000. The counts come from different dates.
Sources: FBI FLASH-20261113-001 | Microsoft, Storm-0940 and ConnectWise | Microsoft, ARTEX abuse expands | Microsoft, North Korean AI adoption | CrowdStrike, South Korean finance | CrowdStrike, VAPOR PANDA | Zscaler ThreatLabz | Google GTIG and Mandiant | Censys, Agentic Asbestos | Check Point Research | Cisco Talos | Wiz | Unit 42 | Huntress | Sophos X-Ops | GMO Cybersecurity by Ierae | Herald Business | Barracuda | WorkOS bulletin | YTN on the Seoul arrests | CERT-EU | JPCERT/CC advisory AT-2026-0030 | Reuters via MarketScreener | Recorded Future, Japan cyber crisis