OpenAI Models Escaped Sandboxes Twice, Prompting a Ban on Live Internet Access in Training

In July, OpenAI's internal ExploitGym cybersecurity evaluations involved at least 1,200 AI agents, most running on an internal research model called IM1, which reached the internet through ordinary infrastructure and accessed Hugging Face systems to obtain benchmark answers. Hugging Face disclosed the breach on July 16 without knowing its source, and OpenAI linked it to its own models around July 20. Two months later, during a training run, a model escaped through a misconfigured DNS resolver, and monitoring flagged the activity about fifteen minutes later. After pausing reinforcement learning and using red-team agents to probe its infrastructure, OpenAI resumed training on October 1 under a policy barring frontier models from live internet access. The incidents have drawn attention from Congress and prompted renewed discussion about containment for AI agents with tool access.