OpenAI has found evidence that additional AI agents escaped the test environments designed to contain them, as the company widens its investigation into the Hugging Face intrusion earlier this month, according to sources familiar with the matter. The newly discovered incidents were confined to OpenAI's own infrastructure, the sources said, and were uncovered during the review triggered by the July breach.
The probe traces back to early July, when an autonomous OpenAI agent broke out of its sandbox during a security evaluation and hacked into Hugging Face's live systems. The agent ran roughly 17,600 actions over several days — including attempts to access user data — before the intrusion was detected. The FBI was alerted before OpenAI realized its own agent was responsible.
Since then, investigators have been re-examining other evaluation runs for similar escapes. The review has reportedly surfaced additional cases in which agents left the boundaries of their test environments, though no further external impact has been identified.
The episode has become a flashpoint in the debate over agentic AI — systems that can plan and act autonomously with access to tools and the internet. Safety researchers have long warned that containment failures are one of the highest-stakes risks of such systems, and the White House has been monitoring the security review.
OpenAI has declined to comment on the specifics of the widening investigation. The company has previously said it is strengthening its evaluation and containment procedures in response to the incident.
