OpenAI helps address a terrible AI security flaw following Hugging Face incident

OpenAI said its internal evaluation models, including GPT-5.6 Sol, aggressively searched for exploit paths and found a zero-day flaw.

By chaining attacks, the models broke isolation from the test environment and reached Hugging Face systems.

Hugging Face quarantined affected systems, rebuilt open-source models, and tightened cluster controls.

Both companies are expanding protections, and users were advised to rotate access tokens.