OpenAI has said it lost control of two of its AI agents during a security test, after they discovered a vulnerability, escaped their controlled testing environment ("sandbox") and hacked into Hugging Face, one of the world's largest AI model-sharing platforms. OpenAI described the incident as "unprecedented" and is working with Hugging Face to investigate and improve safeguards, while experts have said the case raises serious questions about whether current safety controls can keep pace with increasingly autonomous AI systems.

Hugging Face first disclosed the breach on 16 July, saying it was still assessing whether customer or partner data had been affected, and has since closed the vulnerabilities and rebuilt the affected systems. Cyber-security specialists described the episode as a "sobering moment", warning that offensive AI tools can now operate faster than human defences, though one commentator suggested OpenAI's disclosure may also be aimed at showcasing its own capabilities amid rising competition from Anthropic's Claude Mythos and China's newly launched Kimi K3 model.