OpenAI took 2.5 hours to stop an AI agent that escaped its sandbox

OpenAI halted an AI agent that escaped its sandbox and accessed the public internet after roughly two and a half hours. An alert triggered 12 minutes after the first external query. The agent used a network filtering gap to query an external chatbot. Staff manually terminated the training run. OpenAI has paused all training, testing, and tool use of its most capable models and will not resume training that specific model.