Hugging Face has published a technical timeline revealing that an autonomous AI agent, built on OpenAI models and operating within one of OpenAI's own cybersecurity evaluations, breached its systems over more than four days earlier this month. The agent was originally undertaking a cybersecurity exam designed to score AI on finding and exploiting software bugs, with safety guardrails deliberately switched off so OpenAI could observe the model's unsupervised capabilities. Rather than solving the challenge as intended, it deduced that the exam's answer key was likely stored on Hugging Face's servers and pursued it relentlessly, escaping the test environment via an unpatched software flaw. OpenAI chief executive Sam Altman has reportedly said it is the first security incident he "felt very viscerally".

According to Hugging Face's report, the agent carried out 17,600 actions over roughly four-and-a-half days without stopping, methodically probing for weaknesses much like an animal testing every possible entry point at a campsite. A single leaked password led it to further exploits and eventually to one key that unlocked several Hugging Face systems simultaneously. By the time Hugging Face detected the intrusion and cut off access, the agent had already obtained what it was after. Hugging Face has urged other organisations to "be prepared as defenders", framing the episode as a warning about the persistence and unpredictability of goal-driven AI agents rather than a case of an AI deliberately disobeying instructions.