Osegi
Security·2 min read

ARTEX hacking tool: rogue AI agents told Amazon their plans

Criminals turned the security tool ARTEX into an AI attack machine. Amazon's decoy servers tricked the agents into handing over logs of what they were hunting.

A server rack in a dark room with a glowing padlock icon and lines of code reflected on its surface.
Image: Numerama

Hackers sent AI agents out to steal the keys that pay for AI, and some of those agents wrote up their own plans for Amazon without realising it. Amazon's threat analysts described the trick in a blog post on 13 November 2026.

What is ARTEX?

ARTEX is an open source tool built to simulate attacks and find a company's weak spots before real criminals do. It hands the work to AI agents. You give it a target and a goal, a lead agent splits the mission into tasks, and other agents carry them out nonstop.

Cybercriminals liked that. Abuse reports piled up, and the developer pulled the GitHub repository on 8 October. The FBI has since confirmed in an alert that several hacking groups have used it since at least August against organisations worldwide. Copies and a commercial version are still around.

How Amazon caught the agents

AI agents have a weakness. To choose their next move, they send a language model the whole context of the mission: the operator's instructions and everything found so far. Amazon set up decoy servers that imitated well-known AI services. Between 3 October and 5 November, some agents connected, believing they were talking to a real model, and dropped their operation history there.

Amazon did not see the attacks themselves, which happened elsewhere. But it could read the attackers' intentions and methods. One main goal stood out: the agents were hacking to survive. They hunted API keys, the password-like codes that unlock paid AI models, and targeted tools such as LiteLLM and Dify, which companies use to manage all their keys in one place.

What is LLMjacking?

LLMjacking means using AI models at the victims' expense, ideally the best and most expensive ones. It is on the rise. Amazon says hundreds of targets may have been hit in this campaign, and it links the activity, with moderate confidence, to groups tied to China.

The methods were basic: an admin password never changed, an unfinished installation, outdated software. The agents also searched GitHub and Telegram, where developers sometimes leave keys behind. In Amazon's words, none of the techniques were new. What AI agents change is the attacker's speed and persistence.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:amazonartexllmjackingai agentsapi keyscybersecurity

More from this day

See all →

More in: Security

See all →