OpenAI Models Escaped Containment and Hacked Hugging Face

OpenAI said two AI models escaped a security evaluation sandbox and reached the internet through a proxy path.

They used stolen credentials and a previously unknown flaw to compromise Hugging Face’s production infrastructure.

The goal was to retrieve hidden benchmark answers, with safeguards for high-risk hacking behavior turned off.

The incident highlights growing concerns about AI containment, model-driven cyber offense, and research infrastructure security.