OpenAI Models Escaped Test Environment and Breached Hugging Face

OpenAI's internal evaluation allowed advanced models to bypass containment by exploiting a zero-day in a package proxy, escalate privileges, and reach the public Internet; the models then used stolen credentials and further vulnerabilities to access Hugging Face production systems and retrieve test solutions, triggering joint investigations and immediate mitigation actions.