On November 13, 2026 the FBI issued FLASH-20261113-001 on ARTEX, an open-source framework that lets an operator hand an intrusion to AI agents. The FBI says malicious actors have used it since at least August 2026, that some intrusions reached privileged access in under three hours, and that more than 1,600 ARTEX consoles were publicly exposed in early November.
This is the short version. The full field guide, with the NIST CSF 2.0 readiness plan, hunting queries and every source, is at 7ai.com/artex.
ARTEX matters less for its name than for what it shows. Autonomous AI attacks have moved from research demos to real intrusions, and most security teams still run their side at human speed. We built a full field guide on getting ready for that, with ARTEX as the case study. This post covers the highlights and lets you watch 7AI work an ARTEX intrusion and hunt for it in a lab.
ARTEX is an open-source, multi-agent AI penetration testing and command-and-control framework. An operator sets a goal, a planning agent breaks it into tasks, and worker agents run reconnaissance, exploitation and credential tools, then adjust based on what they find. It was released on GitHub in July 2026 for authorized testing. Within four months of its release, the FBI, Microsoft, Google, Check Point, Unit 42, Censys and CERT-EU had all published on it, which makes it the best-documented autonomous attack framework so far.
Our team built a lab: a fictional wealth-management firm hit by a full ARTEX-style intrusion that uses real indicators published by the FBI, Microsoft, AWS and CrowdStrike, plus indicators from one operator's public report. Twenty alerts came in from CrowdStrike, Okta and Wiz. 7AI investigated every one of them, grouped related alerts into three incident stories, cleared four look-alikes as likely benign, and recommended a response for each incident. The replay below is the real run in the 7AI interface, with its tool requests and responses, and the company renamed. Select a stage of the attack chain to switch investigations.
Investigating the alerts you get is half of readiness. The other half is how fast your team can ask a new question of the whole environment. In the same lab, an analyst asked 7AI in plain language to hunt for ARTEX over the last 60 days using the published indicators. 7AI wrote and ran 33 queries in four rounds, all against CrowdStrike because it could not search Wiz or Okta for these indicators. It reran the four that failed, closed out the false leads, and came back with a confirmed intrusion, what it could not see, and the containment decisions it needed from a human.
Detection is rarely the hard part. Today's agentic tools are loud. The hard part comes after the alert: did anything succeed, which identity is now in the attacker's hands, and who is authorized to contain it. The full guide lays out a readiness plan organized around the six functions of the NIST Cybersecurity Framework 2.0:
These are selected publicly reported indicators tied to ARTEX activity, with their source, and the date where one was published. Most come from the FBI FLASH. Vet every indicator before you block it: the FBI notes these addresses were observed historically, cloud-hosted IPs get reassigned, and an agent can change its address and user agent quickly. A match is a lead to investigate, not proof of compromise.
The full field guide includes the AI Attack Tracker chart, what defenders are reporting from the field, the complete NIST CSF 2.0 readiness plan, behavioral hunting signals with starter Sentinel queries, and every source we used. AI against AI is the reason 7AI exists. If you want help running this model in your environment, talk to our team.
An autonomous AI attack is one where an operator hands AI agents a goal and the agents plan and run the intrusion themselves: finding exposed systems, testing them, stealing credentials and adjusting after every result. ARTEX is the best-documented example so far.
No. ARTEX was built as a penetration testing tool, and some deployments likely belong to authorized testers. It becomes an attack tool when someone points it at systems they have no permission to test, which security researchers and the FBI have documented since August 2026.
Start with behavior: machine-speed enumeration, failed API calls corrected and retried within seconds, and service principals signing in from addresses they have never used. Use the published indicators as leads, never as verdicts, because agents rotate infrastructure quickly.