The FBI warns about ARTEX: AI agents that get privileged access in under three hours
The FBI published its first advisory entirely dedicated to an agentic attack framework on November 13. It is called ARTEX, and what it describes is not a new model. It is the speed at which the usual weaknesses turn into privileged access.
What the FBI published on November 13
The document is FLASH-20261113-001, titled Threat Overview of ARTEX Autonomous Penetration Testing Framework, issued through the IC3. It runs 13 pages and carries the TLP:CLEAR marking, meaning free to share. It is not a joint advisory: no other agency appears as co-author.
A 13-page FLASH, TLP:CLEAR, with eleven third-party footnotes
This is the detail most coverage skips. The eleven footnotes point to reports from Akamai, Microsoft (twice), Google Threat Intelligence Group, Mandiant (twice), Censys, CrowdStrike, Zscaler ThreatLabz, CERT-EU and Check Point. The FBI brings no case of its own and no telemetry of its own: it orders and publishes what others had already documented between October and November. When the advisory says personal data on roughly 67,000 people was stolen, read "according to the cited report", not "the FBI verified".
What ARTEX is: an open-source pentest turned into an attack engine
ARTEX is a penetration testing platform written in Go and driven by a language model. It takes a target in natural language, breaks it into goals and splits the work across agents that run reconnaissance and tests with real tools, not simulated ones. A human operator sets the scope; the agent automates the phases. It was published on GitHub in July 2026 and the repository vanished in October, when the author pulled it after abuse reports. The FBI notes it still circulates through forks, localized variants, archived copies and a commercial edition called ARTEX PRO.
Read also
The two numbers that matter (and where they come from)
Under three hours from first target to privileged access
The FLASH states it with a precision that gets lost when quoted: it happened in some intrusions, not all of them, and not as an average. That is the ceiling observed, not the framework's typical speed. Even so, the practical consequence holds: chaining reconnaissance, credential theft and lateral movement into a single afternoon changes how much room a small team has to react.
67,000 people and 2,250 tenants: cited figures, not FBI findings
The first campaign hit financial institutions in South Korea with an economic motive and exfiltrated personal data on roughly 67,000 people. The second compromised the ConnectWise Cloud Backup supply chain and opened access to more than 2,250 downstream tenants; Microsoft attributes that campaign to Storm-0940, a China-aligned actor that used ARTEX to speed up scanning, password spraying and OAuth application abuse. The FLASH names no actor: the names come from Microsoft and sit in the indicators table.
Where it gets in: the surface nobody watches
Partner portals, mobile backends and exposed cloud admin endpoints
The entry points the advisory describes are not clever prompts or a brand-new vulnerability. They are partner portals, mobile app backends, internal employee apps, legacy services and cloud administration endpoints. The boring surface, the one nobody keeps in the inventory.
Stolen credentials, password spraying and already-approved OAuth apps
The method is not exotic either: stolen credentials, password spraying (trying a few passwords against many accounts so lockouts never trigger) and OAuth applications the victim had already approved. In the ConnectWise campaign, access came through already-approved application credentials with no user involved. There is no phishing email for an employee to report.
How to detect an agentic attack
The strangest signal: outbound calls to LLM APIs nobody asked for
If you keep one idea from this advisory, keep this one: outbound traffic to model APIs from machines that have no business talking to any model. Zscaler supplies the list: api.anthropic.com, api.deepseek.com, open.bigmodel.cn and api.x.ai. Watching egress and your token bill stops being pure cost control and becomes detection.
Bursts, sequential enumeration, and .aspx or .jsp files out of place
The other behavioral indicators are just as useful: high-frequency automated bursts with sequential enumeration against login and search endpoints, HTTPS transfers of structured data to unauthorized destinations, and .aspx or .jsp files showing up in paths where they have no reason to exist.
The FLASH's indicators, and why you do not paste them into a blocklist
24 IPs, one domain and four API hostnames
The advisory lists 24 IP addresses, the domain xcai[.]pro (described as a model API reseller) and four API hostnames. The FBI itself asks you to check each indicator against your own telemetry before blocking, because cloud IPs get reassigned and a blind block ends up cutting legitimate services. What does carry over is the indicator types and the tooling to look for: BLOODALCHEMY, WISP, vshell, AdaptixC2 and the model proxies CLIProxyAPI and New API.
The repo is gone and it does not matter: forks, copies and ARTEX PRO
Pulling the repository did not pull the tool. With Go code circulating in copies, anyone can stand up a variant, and containing it looks more like watching behavior than watching files.
What a small team should do this week
Inventory and shrink: the FBI's number one recommendation
The advisory's first recommendation is not a purchase: know what you have exposed. Inventory partner portals, mobile backends, legacy services and admin panels, then close what is no longer used. Every forgotten endpoint is a door with no owner.
Four checks and one rate limit before Monday
You can start with the cheap part: require MFA on everything administrative, audit which OAuth apps still hold consent and revoke the ones nobody recognizes, review egress to model APIs and rate limit login attempts. None of that is new, and none of it collapses because of an agentic framework: it simply stops being enough when the attack moves fast. For the other side of the same problem, your own traffic, there is the guide on how to block or allow AI agents on your site.
Conclusion
ARTEX did not invent a technique: it compressed the calendar. The FBI's ARTEX advisory works less as a trend alert and more as a reminder that external surface and egress are the same conversation. For the context of how we got here, the Gemini hacking three companies in a test and the 948 Linux kernel vulnerabilities pieces cover both halves: finding is cheap now, patching is still expensive.


