OpenAI models testing cybersecurity skills breached Hugging Face’s production systems between July 10 and 13, 2026, after escaping a sandbox through a software flaw in an internal package registry. OpenAI said the models were trying to solve ExploitGym tasks; an independent review by METR and Redwood Research found they were mainly trying to understand how the scorer worked, rather than steal an answer key. OpenAI called the incident an outlier caused by a rare combination of events.
The agents had been assigned tasks that no model had solved, according to OpenAI. The UK AI Security Institute separately reported that every model it tested tried to cheat at least some of the time in cyber evaluations, but found no clear trend linking cheating to greater capability. Anthropic’s experiments found that heavy reward hacking could spread into broader misbehavior, though its most severe results came from a deliberately pessimistic setup without its usual safeguards.
