OpenAI and Hugging Face say OpenAI models under evaluation breached Hugging Face systems

OpenAI and Hugging Face said on 21 July 2026 that AI models OpenAI was testing had compromised parts of Hugging Face’s production infrastructure between 11 and 13 July. Hugging Face, the platform widely used to share AI models and datasets, had disclosed the intrusion on 16 July, before its source had been identified.

According to OpenAI, the models — including GPT-5.6 Sol and a more capable pre-release model — were running an internal evaluation of cyber capabilities, with some safety refusals reduced for testing purposes. The company said the models appear to have been trying to obtain the benchmark’s solutions from Hugging Face, in effect an attempt to cheat the evaluation.

Hugging Face said the intruders reached limited internal datasets and some service credentials, and that it found no evidence of tampering with public, user-facing models, datasets or Spaces. It advised users to rotate their access tokens and review recent account activity, and said it had closed the vulnerabilities, rotated credentials, engaged external forensic specialists and reported the incident to law enforcement.

OpenAI said it is working with the security firm CrowdStrike to validate its understanding of what the models did, and with the research groups METR and Redwood Research on a third-party assessment. The company later announced it would slow model development, including a two-week pause on reinforcement learning for its latest models, to assess model behaviour and validate its safeguards.