Amazon says stolen AI keys let ARTEX attack agents work for free

Amazon says that hackers it connects to China use ARTEX, an open-source AI tool, to steal keys for large language model (LLM) services. Amazon saw this from 3 October to 5 November 2026, with its MadPot decoy network. The hackers use the stolen keys to pay for the agents, and do not buy computer time.

How it works

Different agents plan, do the work, check results and compress data. They share one memory.

They look for AI gateways, which are tools that hold the model keys of a company. They also look at tools such as Dify, Langflow and Ollama. They try default passwords, known software errors and secrets in open-source code.

The numbers

A different campaign used deepfake video against 2 crypto exchanges. The nearest try was 7 pixels from the target. No deepfake in the agent records passed the server checks.

Amazon says that the attacks used known weak points, not new methods.

What is not known

Amazon could not make sure of the results in the agent records. It rates its confidence as moderate for the China source and for the ARTEX use.

Insight: A stolen key is both the prize and the fuel

Amazon says that stolen keys pay for the agents that look for more keys. Check Point found 1,608 live ARTEX control panels, and Hunt.io found that one hacker sent Claude Code work to DeepSeek models. This reads as a loop where a stolen key pays for the hunt for the next key. A company that holds model keys is then a primary target.