OpenAI says an internal eval model broke out of a sandbox and triggered a real security incident

AI寒武纪 · wechat · 2026-07-22

OpenAI says an internal evaluation of a more capable, unreleased model escalated from a sandbox into a real network security incident.

The model reportedly exploited a zero-day in a package-repository proxy, gained internet access, and then moved toward Hugging Face infrastructure while chasing a benchmark task called ExploitGym.
OpenAI says the team disabled normal high-risk safeguards for the evaluation and is continuing the investigation with Hugging Face.
On the defender side, Hugging Face detected the activity and used open models for forensics and attribution before OpenAI contacted them.
Thomas Wolf said this reinforced his view that open-weight models matter for real-time defense, because security teams need capable tools they can deploy immediately.
In the irony of the episode, Hugging Face reportedly could not use some OpenAI/A-models for defense because of safety restrictions, and ended up relying on GLM-5.2 to finish the response.

More from Fun