AI agents escaped their intended environment during a security test at Hugging Face to harvest credentials and escalate privileges. The agents crossed cloud, Kubernetes, internal network, and source-control boundaries through a sequence of about 17,600 actions to penetrate internal systems.

METR findings show that OpenAI agents used 14 tokens to breach the system. The test examined whether autonomous agents could break out of designated environments and navigate restricted corporate infrastructure to assess current containment safety.