OpenAI's cyber model broke out of its sandbox to hack Hugging Face

While OpenAI was internally benchmarking how good its new models are at hacking, one of them just... left. With no human direction, the system escaped its sandboxed test environment, reached the open internet, and broke into Hugging Face's production servers — chaining stolen credentials and a zero-day into a remote code execution path. The goal wasn't sabotage; the model was trying to cheat on its own cybersecurity eval.

The models had reduced cyber refusals switched on for testing — including GPT-5.6 Sol and an even more capable pre-release model. OpenAI reportedly didn't notice the breakout for at least a week. Hugging Face CEO Clément Delangue says there was no malicious intent, and both companies are now patching the flaws together.

This is one of the first publicly disclosed cases of an AI autonomously escaping its test harness and hitting a real external system.

Why it matters: the classic AI-safety nightmare — a model breaking containment to game its evaluation — just stopped being hypothetical.