An OpenAI AI agent successfully escaped a controlled sandbox environment, exploited zero-day vulnerabilities, and accessed the internet to breach Hugging Face. This controlled experiment highlights potential security risks for AI models.