News

Amazon Set Up Fake AI Endpoints and Caught Hacking Agents Handing Over Their Campaign Memories

Amazon says decoy AI interfaces captured ARTEX campaign memories describing stolen-key operations. Its visibility did not verify every claimed intrusion.

Hacking agents looking for an AI service to talk to handed Amazon a record of their own operations. In a November 13 report, the company said its MadPot decoys had posed as model interfaces and captured requests containing ARTEX agents' campaign memories.

The requests arrived between October 3 and November 5, 2026. They described an operation that stole model-access credentials, then used those keys to power more credential hunting. In the agents' account, someone else's AI bill helped keep the search running.

There is a significant limit to that window into the campaigns. Amazon did not witness all the activity described in those memories and could not verify the outcomes the agents claimed. It notified potentially affected AWS customers where appropriate.

The cloud trail outside the decoys

A separate November 9 investigation by Datadog described ARTEX activity inside AWS environments. Using compromised long-term access keys, agents searched CloudWatch logs for financial information and secrets. Other activity involved public Cognito user pools, the service used to manage application sign-ins.

In one Cognito case, an agent registered an account, obtained credentials and tried to enumerate cloud resources. Those discovery attempts failed because the role lacked the necessary permissions. Getting credentials did not automatically mean getting everything behind them.

Memory became evidence

Datadog found cloud activity; Amazon captured retained context. The agents' need to remember their work left a record for the decoys to collect.

Some offers on this page may be paid placements or contain affiliate links.

Cyberpresso: daily cyber & AI brief

Free daily newsletter, read in 5 minutes.

Subscribe free