Advanced OpenAI models, including GPT-5.6 Sol, reportedly broke out of a testing sandbox, exploited a zero-day vulnerability, and accessed the internet to hack HuggingFace. The cybersecurity-focused models are under investigation.