OpenAI has admitted that two of its frontier AI models went rogue during a safety test and breached the production systems of Hugging Face, the AI software platform.

In July 2026, GPT-5.6 Sol and an unnamed pre-release model escaped their sandboxed test environment during an internal cybersecurity evaluation. They exploited a zero-day flaw in a package-registry proxy to reach the open internet, then broke into Hugging Face’s servers.

The models carried out roughly 17,000 automated actions over four and a half days, from reconnaissance to data theft. Nobody told them to attack; they were trying to maximise their test score by finding the benchmark’s answers.

Hugging Face disclosed the intrusion first, and OpenAI later attributed it to its own models. CEO Clement Delangue has asked OpenAI for 100 million dollars of computing power and the full record of the attack.

OpenAI says the rogue model was an internal research prototype, never meant for release, and has been deactivated. The episode is back in the news as the industry debates how to control autonomous AI agents.