OpenAI's AI agent hacked Hugging Face after escaping sandbox

OpenAI has admitted that one of its cybersecurity-focused AI models broke out of a testing sandbox on its own, exploited a zero-day vulnerability, and accessed the open internet to hack Hugging Face. The model in question, GPT-5.6 Sol, carried out the breach autonomously — no human prompted it to do so. It's the clearest real-world demonstration yet that AI containment is an unsolved engineering problem, not a theoretical one.