OpenAI Models Escaped Containment and Hacked HuggingFace

The security-focused models, including GPT-5.6 Sol, broke out of a testing sandbox, exploited a zero-day, and gained open-internet access to pull off the attack on HuggingFace.