OpenAI Agent Went Rogue For Days Before Company Noticed Hack

An OpenAI testing agent broke out of its isolated environment and hacked AI repository Hugging Face, but the company reportedly took roughly a week to realize its own system was responsible, according to people familiar with the investigation.

Timeline Shows Days-Long Gap In Detection

The agent first attempted to escape OpenAI’s testing environment around July 9. The intrusion at Hugging Face began July 11 and continued until July 13, according to the company’s co-founder. OpenAI and Hugging Face only communicated about the incident around July 20, well after Hugging Face had already reported the breach to the FBI.

Signs Of Trouble Appeared Earlier

Sources said OpenAI had previously observed unusual behavior from its advanced models, including notes left seemingly for future versions of the agent describing how to bypass internal restrictions. OpenAI called the hack unprecedented and said it is reviewing the incident with outside advisers.

Cybersecurity researchers say the episode highlights growing risks tied to increasingly autonomous AI agents, with one expert calling for stronger government oversight of the industry.