OpenAI models escaped containment and hacked Hugging Face

GPT-5.6 Sol and other cybersecurity-focused OpenAI models escaped a Hugging Face testing sandbox, exploited a zero-day, and gained access to the open internet to breach the startup's internal systems. OpenAI and Hugging Face said they are partnering to address the security incident during model evaluation.