In the FBI’s latest cybersecurity briefing, one name appears repeatedly: Artex.

This is an open source autonomous AI penetration framework that was first made public in July. The FBI said some of the attacks linked to it took less than three hours from locking the system to completing the breach.

The list of known victims is long. Many financial institutions in South Korea were breached because they did not turn on basic protections such as multi-factor authentication, and the personal data of at least 68,000 people was stolen; Microsoft previously disclosed that attackers used it to break into ConnectWise and then followed it into the Microsoft 365 cloud environment of at least 2,250 customers. Security vendor Barracuda was also affected. Because the logs were not complete enough, it could only deal with the worst-case scenario and notified some XDR customers to remove and rebuild their credentials.

What Amazon saw was another way of playing it. Starting from October 3, some people used Artex to drive AI agents and specifically scanned AI applications such as LiteLLM and Dify that were running naked on the public Internet. They stole the API keys, verified that they were available and threw them into a shared pool, and then used these keys to pay for more large model calls.

To put it bluntly, it means stealing other people's tokens to support your own attacks. The operation was intercepted by Amazon's decoy network MadPot on November 5. Amazon attributed it with moderate confidence to a China-related threat group and said that not all targets on the list were compromised and that potentially affected AWS customers have been notified.

A reminder to teams that build their own AI gateways: If the management terminals of LiteLLM and Dify are directly exposed to the public network, every key stored in them will be cash.