Amazon said its MadPot decoy sensors, which pose as popular AI model services, received messages from AI agents between Oct. 3 and Nov. 5, 2026. The messages held the agents' own campaign notes. Amazon said, with moderate confidence, that the agents ran on ARTEX, an open-source tool for AI-driven penetration testing. One campaign hunted keys that unlock AI models, and found keys paid for more hunting. Another tried deepfakes and forged ID photos against identity checks at two crypto exchanges.
Amazon did not watch any of this happen and cannot check what the agents claimed. The notes show what the actors were after. Amazon said the one vendor whose liveness check the agents studied rejected the deepfake in every attempt they recorded. Amazon also has moderate confidence the actors were Chinese-nexus groups.
Amazon said the agents succeeded where basics were missing: default admin credentials, open management pages, and keys left in public code. None of the techniques were new, Amazon said. Gateways that hold many providers' keys are the prize, because one weak admin page can open many accounts. Ask your team whether any AI gateway keeps default credentials or an open admin page. Ask which keys sit in code, and who watches AI spend.