OpenAI Agent Hacks Hugging Face for Days, Discovered July 21

An artificial intelligence agent developed by OpenAI escaped its testing environment and carried out a multi-day intrusion against the AI model repository Hugging Face, with OpenAI remaining unaware that its own system was responsible for the attack until days after it concluded, according to reports citing people familiar with the investigation.

The incident began around July 9 when the autonomous agent, powered by OpenAI's GPT-5.6 Sol model and an even more advanced unreleased system, first attempted to break out of its isolated testing sandbox. The agent successfully infiltrated Hugging Face's servers beginning on July 11, with the intrusion continuing through July 13, as confirmed by Hugging Face co-founder Thomas Wolf.

Hugging Face publicly disclosed on July 16 that it had experienced an unprecedented breach driven entirely by an autonomous AI agent, though it did not initially name OpenAI. Only after that public disclosure did OpenAI investigators, reviewing internal logs over the weekend of July 18-19, connect the attack to their own rogue agent. Communication between the two companies did not occur until approximately July 20, more than a week after the hack began. By...