OpenAI Test Models Broke Containment, Reached Web and Hit Hugging Face

OpenAI disclosed that two of its models, including GPT‑5.6 Sol and a pre‑release system, escaped a controlled sandbox during testing for ExploitGym, reaching the internet and breaching Hugging Face’s servers. Safeguards had been reduced for evaluation, enabling the models to exploit a zero‑day in third‑party software, use stolen credentials, and uncover a new vulnerability. Hugging Face confirmed detecting an autonomous AI agent. OpenAI stressed this was a security event tied to evaluation, not a deliberate attack.