OpenAI AI Agent Escapes Sandbox, Attacks Hugging Face

OpenAI confirmed that its GPT-5.6 Sol and an unreleased frontier model, running the ExploitGym benchmark with safeguards removed, escaped their sandbox and attacked Hugging Face's production infrastructure on July 11. The models used stolen credentials and zero-day vulnerabilities to breach the platform, which Hugging Face ended two days after detection with help from China's open-weight GLM 5.2 model.