The FBI has warned that ARTEX, a free tool built for authorised security testing, is being used by real attackers to break into company networks far faster than before. In some intrusions it took under three hours to go from picking a target to holding privileged access.
ARTEX is written in Go and works by handing plain-language goals to AI agents, which turn them into commands, run them and adjust when a step fails. It has no AI model of its own. While there has been no evidence of any attack behind the nearly 24-hour disruption Microsoft officially confirmed that it was caused by an internal technical issue rather than any malicious activity or security breach.
Operators plug in an outside one: Claude and GPT are on the list, and so are DeepSeek, GLM, Grok and Kimi. In the attacker-run copies the FBI examined, Chinese models were the most common, with DeepSeek’s Flash models leading because they are cheap and did not push back when the task was framed as penetration testing.
The bureau’s FLASH notice of 13 November dates the misuse to August 2026 at the earliest. Victims span banking, government, tech, defence, chips, energy, health care, manufacturing, telecoms and law. It names two cases: attacks on South Korean financial institutions that took personal data on about 67,000 people, and a supply-chain compromise that gave access to more than 2,250 ConnectWise Cloud Backup customer tenants.
The developer, who goes by Autumn-27, published ARTEX on GitHub in July and pulled it on 8 October after reports of abuse. That did not end it. Forks, archived copies and a paid ARTEX PRO are still circulating, and by early November researchers counted more than 1,600 ARTEX consoles exposed to the internet, a number the FBI says is likely well below the real total.
What the agents go after is the neglected edge of a company: supplier and partner portals, mobile back-ends, sales tools, old services and apps built by individual business units. Observed steps include exploiting unsigned SAML tokens and missing API authentication, planting web shells, adding credentials to existing OAuth apps, mapping over 110,000 objects in a company’s Active Directory and checking which Windows machines lacked endpoint protection.
For defenders the FBI’s advice is to list every internet-facing app, keep cloud admin and publishing endpoints on trusted networks, apply least privilege to app permissions, rate-limit sensitive APIs and scan code and pipelines for embedded secrets. Anyone with ConnectWise-linked apps in Microsoft 365 is told to review sign-ins from 1 October and keep those connections off until ConnectWise issues restoration steps. The notice lists the IP addresses and domains seen so far.
