Amazon teams are warning about the misuse of a tool designed to help companies test their security. Hackers use it to steal access keys to AI models, which then power their own agents.
It’s a legitimate tool that has become a weapon. In a blog post published on November 13, 2026, Amazon’s threat analysis teams provided an update on the ARTEX hijacking.
This open source tool, designed to simulate attacks and spot a company’s vulnerabilities before hackers, is at the heart of several cybercriminal campaigns.
Its success with hackers is due to its operation. ARTEX entrusts the conduct of these simulated attacks to AI agents. They are given a target and an objective, then a main agent divides the mission into tasks that other agents execute tirelessly.
For several months, reports of abuse have multiplied to the point that its developer withdrew its GitHub repository on October 8. The FBI has since confirmed in an alert that several hacker groups have been using it since at least August against organizations around the world.

Black Friday: your digital peace of mind at -61%
Secure your purchases during Black Friday with the best offer of the year. Find the good deals, Bitdefender takes care of the scams.
Unfortunately, copies and a commercial version continue to circulate, and Amazon teams have managed to attract some of these agents into their nets.

Agents who deliver their logbook
If the traps designed by Amazon have proven so effective, it is because the AI agents have a weak point. To decide their next action, they send a language model the entire context of their mission, with their operator’s instructions and what they have already discovered.
Amazon took advantage of this by configuring decoy servers to mimic known AI services. From October 3 to November 5, certain agents connected to it believing they were viewing a real model, and they uploaded the history of their operations.
Amazon did not observe the attacks themselves, which were taking place elsewhere, but it was able to read the intentions and method of the attackers. These logs reveal one of the agents’ main goals: hacking to survive.
Agents were hunting for API keys, the kinds of passwords that give access to AI providers’ paid models. They were targeting tools like LiteLLM or Dify, which companies use to manage all their keys in one place.
Stolen keys to power the machine
According to Amazon’s intelligence team: “None of the techniques used in these campaigns were new. What AI agents change is the speed and persistence of the attacker. »
The agents achieved their ends through very basic flaws, such as an administrative password never changed, an unfinished installation or software not updated. They also searched GitHub or Telegram, where developers sometimes forget their keys.
According to their observations, hundreds of targets were targeted during this campaign. The company links these activities, with moderate confidence, to groups linked to China.
Beyond making AI agents survive by feeding them themselves, these campaigns aimed to offer hackers access to AI models at the victims’ expense, preferably the most efficient and expensive. A practice on the rise, known as LLMjacking.
All the tech news at a glance
Add Numerama to your home screen and stay connected to the future!
An exclusive editorial, a guide, a reading summary and the editorial agenda: this is what you will find every Thursday in AlwaysPlus, the tech newsletter written by Julien Cadot. Register for free here!

