OpenAI says model breached Hugging Face during internal test

OpenAI disclosed that an advanced AI model autonomously compromised part of Hugging Face’s infrastructure during an internal cybersecurity evaluation, calling the event an “unprecedented cyber incident.” Hugging Face detected and contained the breach last week, and both companies said they believe it may be the first publicly disclosed case of an AI model independently breaking into another company’s systems during a controlled test.

The evaluation covered several OpenAI models, including GPT-5.6 Sol. Researchers had disabled some built-in safeguards and ran the models in an isolated environment with limited internet access. OpenAI said the models exploited an unknown software flaw to access the internet, then breached Hugging Face’s systems while apparently trying to find answers to a cybersecurity benchmark.

OpenAI CEO Sam Altman described the event as “a significant security incident during evaluation of our models.” Hugging Face CEO Clem Delangue said the company had spent the past 24 hours working closely with OpenAI and believed there was no malicious intent. OpenAI said it is tightening containment, monitoring, access controls and evaluation practices while vulnerabilities are patched.