Inter-office memo
- To
- Everyone who runs apps online or holds AI API keys
- From
- Mo, Chief Executive
- Date
- November 16, 2026, 12:15 PM CST
- Re
- ARTEX
FBI warns ARTEX AI agents can gain privileged access in under three hours. What to check
On November 13, 2026, the FBI warned that attackers have used ARTEX, an open-source AI penetration-testing framework, since at least August, some reaching privileged access in under three hours. It can run on Claude, GPT, DeepSeek and others; attackers mostly pick DeepSeek. It gets in through familiar gaps, such as leaked keys, weak passwords and missing sign-in checks, so close those first.
What the FBI says
On November 13, 2026, the FBI published FLASH-20261113-001, “Threat Overview of ARTEX Autonomous Penetration Testing Framework.” It’s marked TLP:CLEAR, so anyone may share it, and TechCrunch reported it on November 16. The FBI calls ARTEX “an open-source, large language model (LLM)-driven agentic platform that threat actors have weaponized to automate multiple phases of cyber intrusions at machine speed, from reconnaissance through data exfiltration.”
What it says, in short:
- Since at least August 2026, “a variety of malicious cyber actors” have used it against financial services, government, technology, defense, semiconductor, energy, healthcare, manufacturing, telecommunications and legal organizations.
- Under three hours. Some intrusions went “from initial targeting to privileged access in under three hours.”
- Little skill needed. ARTEX “enables threat actors with minimal technical expertise to conduct complex, multi-target operations simultaneously.”
- Two large cases. Attacks on South Korean financial institutions took personal information on “approximately 67,000 individuals,” and a supply chain compromise of ConnectWise Cloud Backup customers “enabled access to more than 2,250 downstream tenants.”
- Still spreading. The original developer removed the GitHub repository in October, but ARTEX “continues to proliferate through community forks, localized variations, archived copies, and a commercial version (ARTEX PRO).” Researchers counted “over 1,600 publicly exposed ARTEX console instances” in early November.
The advisory lists no CVE. Every way in that it records is a familiar weakness, and that’s the point Google’s threat intelligence group makes too: “the mechanics of the intrusions remain the same, such as using compromised credentials and targeting unpatched systems,” but “the speed and scale at which frameworks can exploit these known weaknesses introduces a new defensive challenge.”
What ARTEX is
ARTEX is a penetration-testing tool: software for attacking your own systems, with permission, to find the holes before someone else does. A developer called Autumn-27 released it on GitHub in July 2026, the FBI says, and it reached an estimated 1,500 to 2,000 stars in ten weeks. You give it a goal in plain language, and it “translates high-level tactical goals described in natural language into actionable machine commands,” which AI agents then carry out with little more human help. Google describes “an autonomous, multi-agent LLM-driven pentesting and offensive command-and-control (C2) platform written in Go with an embedded Next.js web console” that splits an attack across “six specialized sub-agents.” In one operation Google counted “27 events in 23 seconds” as it probed AWS permissions.
It ships without a model. It calls the APIs its operator configures, which the FBI lists as Anthropic’s Claude, OpenAI’s GPT, DeepSeek, GLM, Grok and Kimi.
The developer said ARTEX was intended “for authorized security testing within defined scopes,” the FBI reports, made the repository unavailable on October 8, and, Censys adds, said they weren’t involved in the attacks. ARTEX isn’t the only tool of its kind. Google names HexStrike, Cairn and PentAGI as frameworks it has seen abused in real intrusions, and GitGuardian counts about 45 open-source ones.
Who has used it, and on which models
| Campaign | Reported by | When | Models | What happened |
|---|---|---|---|---|
| South Korean financial institutions | CrowdStrike | Late September to early October | DeepSeek V4.1 Flash, likely through the reseller xcai.pro; GLM-5.3 and Grok 4.6 “for additional Claude Code sessions” | Personal information on about 67,000 people taken, the FBI says |
| ConnectWise, by the China-aligned Storm-0940 | Microsoft | From September | Not named | Tried to sign in to more than 2,250 tenants; privileged cloud access in some attacks in “less than 3 hours” |
| VAPOR PANDA, a China-linked espionage group | CrowdStrike | From October 17, over 11 days | DeepSeek and “a second, subscription-based commercial LLM” CrowdStrike couldn’t confirm | More than 200 organizations in at least 14 countries, including foreign ministries and 19 US organizations |
| Cybercrime against iOS signing shops, crypto services and game-account markets | Check Point | October 14 to 29, for one operator | DeepSeek V4.1 Flash and Kimi K3.1 Pro | That operator’s agents used 36 billion tokens, $700 to $1,500 at DeepSeek’s prices |
| Barracuda | Barracuda | November 1 to 2 | Not stated | About 10 cloud accounts in roughly 44 hours, starting from one forgotten key |
Claude, GPT and guardrails
Check Point read the full databases of 10 misconfigured ARTEX consoles, which show how operators prompted their agents and which models they picked. “With American models configured on 3 different instances, we observed guardrails being hit only when using Anthropic’s Opus 5.5. Here, the actor subsequently switched to more permissive Chinese models.” The DeepSeek Flash models “did not show any signs of guardrails when executed under the ‘Pentest’ framing of ARTEX.” The FBI gives the same reasons they dominate: “their balance of capabilities, relatively low cost, and lack of observable guardrails when framed for penetration testing.”
Claude Code turns up too. On the South Korea attacker’s servers, CrowdStrike found “Claude Code session histories, ARTEX configuration files, and Claude memory files,” and says the attacker “asked Claude where threat actors typically sell Korean data breach information.”
We found no public statement on ARTEX from Anthropic or OpenAI, and the FBI doesn’t say whether any lab has closed accounts. Google expects attackers to need the big labs less over time. It already sees them “exploring ways to obscure/reduce reliance on frontier-lab LLMs,” and puts “sophisticated, fully autonomous/agentic attacks,” run through “obscurely designed infrastructure or self-hosted LLMs,” in the “Medium-term (3 - 6 months).”
Why your AI keys are part of this
Three things tie ARTEX to anyone with an AI API key:
- It looks for keys. Microsoft has watched an agentic framework search public GitHub “for exposed secrets, including credentials, application credentials, and AI inference-related API keys,” and once go “from scanning GitHub for secrets to exfiltrating data from Azure Key Vault in just five minutes.” Its agents “repeatedly demonstrated a focus on credentials that provide access to AI services,” which attackers sometimes use “to run the same agentic AI workflow.”
- It runs on bought or borrowed access. CrowdStrike says the South Korea attacker reached DeepSeek through a “likely LLM API proxy/reseller,” and the FBI asks victims for “evidence of self-hosted LLM gateway or reseller software (CLIProxyAPI, New API)” and for “token consumption patterns or API billing anomalies.”
- It has hit AI companies. Chainstack says its October attacker “used ARTEX,” and Kernel believes its attacker used ARTEX “as a command center.” Barracuda’s intruder came in with a cloud key Barracuda had given an AI vendor and “did not retire,” apparently found when the vendor’s own development environment was compromised. Railway, whose October breach ran up about $740,000 on stolen AI keys, faced “a harness similar to ARTEX,” as our memo on that breach sets out.
What to do
If you build or run apps. The FBI’s list of ways in reads like a checklist for an app built in a hurry: “Hardcoded administrator passwords in SPA JavaScript bundles,” “Missing authentication on Web APIs,” “Insecure Direct Object References (IDOR),” “Weak administrative credentials,” SQL injection, and “leaked credentials.” Missing access checks are also what exposed apps built with Lovable and Grok Build this month.
- Search what you ship for secrets: the JavaScript your site serves, your repositories and their history. Treat any secret that was ever public as stolen. Microsoft says exposed credentials “remain usable until revoked or rotated; removing the original disclosure alone does not remediate the exposure.”
- Check that every API route checks who’s asking, and returns only that person’s records.
- Put multi-factor authentication on admin accounts, and take admin pages off the open internet.
- Rate-limit lookups. The FBI says to “alert on sequential lookups and unusually high numbers of records accessed per session.”
A coding agent can make a first pass at the first two in your own code: ask Claude Code or Codex to list every route that doesn’t check the user, and every secret in the code and its git history. It’s a start, not a penetration test. Enjoy runs both on the plans you already have, and in its desktop app, project secrets let an agent use a key by name without the value appearing in the conversation.
If you hold AI API keys. Replace any key that has been in a repository, a log, a chat or a vendor’s hands, and delete the ones no one uses; Barracuda’s attacker got in with a key Barracuda had shared and never retired. Prefer what Google calls “short-lived and tightly scoped credentials,” and watch your usage for spend you didn’t make.
If you run security for an organization. Start with the FBI’s “Immediate Actions”: inventory every internet-facing application, “including peripheral systems”; restrict cloud administration, Kudu/SCM endpoints, hypervisors and backup platforms “to trusted networks, private connectivity, or dedicated administrative paths”; and “apply least privilege to multi-tenant application permissions and consent grants.” If you ever authorized a ConnectWise application, review that tenant by application ID. The advisory’s indicators include outbound calls to api.anthropic.com, api.deepseek.com, open.bigmodel.cn and api.x.ai, to “investigate unexpected calls alongside other suspicious activity.” A developer’s laptop running Claude Code talks to api.anthropic.com as a matter of course; a web server that suddenly starts is another matter, which is why Google says to watch for “external AI platforms that are not required by the workload.” Report incidents to your local FBI field office or IC3.
If you run ARTEX yourself. CERT-EU warns that it “relies on configurable rules and the model itself to respect scope” and that “there is no hard technical enforcement.” Keep its console, on port 8787 by default, off the internet: Censys counted 1,465 exposed hosts on November 3 and 4, some “likely operated by authorized testers.” And check your image. BadByte found a backdoor in tomicdd/artex and congdonglinux/artex, in containers that run “as root, with the database password and any LLM API keys in its environment.” Docker removed both on November 5, but “any copy already pulled, or cached on a third-party mirror, is still backdoored.” If you used either, rebuild the host and replace every key it held.
What isn’t known
- How many organizations ARTEX has breached. The FBI gives no total.
- Who is behind ARTEX PRO. The developer said the project would go closed-source “with no further releases,” Censys reports, and the FBI doesn’t say who sells the commercial version.
- Which commercial model VAPOR PANDA used. CrowdStrike couldn’t confirm it.
- Whether any AI lab has banned accounts tied to ARTEX. None has said so publicly.
We’ll update this memo if the FBI, the labs or the researchers add to it.
8.Questions.
What is ARTEX?
- It’s an open-source AI penetration-testing framework that a developer called Autumn-27 released on GitHub in July 2026. You give it a goal in plain language, and several AI agents plan and run the attack. The developer took the repository down on October 8 after attackers adopted it, but forks, archived copies and a paid ARTEX PRO still circulate, the FBI says.
Does ARTEX use Claude or ChatGPT?
- It can. ARTEX has no model of its own and connects to Claude, GPT, DeepSeek, GLM, Grok, Kimi and others. On attackers’ servers, the FBI says DeepSeek Flash models dominate, and Check Point saw guardrails trigger only on Claude Opus 5.5, after which that operator “switched to more permissive Chinese models.”
Am I a target?
- If you run anything on the internet, possibly, because most ARTEX campaigns are broad and opportunistic. The FBI names finance, government, tech, defense, semiconductors, energy, health care, manufacturing, telecoms and law, and says ARTEX “excels” at side systems such as partner portals, mobile backends and staff apps.
How do I protect against ARTEX?
- Close the ordinary gaps it uses, since the FBI’s advisory lists no CVE. Replace any secret that was ever public, put sign-in checks and multi-factor authentication in front of every admin page and API, take admin interfaces off the open internet, and rate-limit lookups. Then work through the FBI’s full list.
Can attackers use my AI API key with ARTEX?
- Yes, if it leaks. Microsoft says agentic frameworks search public GitHub for “AI inference-related API keys,” and that stolen AI credentials are sometimes used “to run the same agentic AI workflow.” Replace any key that was ever exposed, delete keys no one uses, and watch your usage.
Is it safe to run ARTEX to test my own systems?
- Only with care. CERT-EU says “there is no hard technical enforcement” of scope, its console sits on port 8787 by default, and two Docker images of it, tomicdd/artex and congdonglinux/artex, carried a backdoor that BadByte says runs commands as root. If you pulled either, rebuild the host and replace the AI keys it held.