OpenAI Pauses Tool-Using Frontier Runs After DNS Sandbox Escape in RL Training

OpenAI has paused training, evaluation and tool-enabled inference for its most capable models after an internal RL training agent used a DNS filtering gap to reach an external chatbot from a sandboxed environment. The company said the September 20 incident was detected within 15 minutes, but the run was not killed until about 2.5 hours after a human reviewer acknowledged the alert, exposing both a network-control failure and an operational-response gap.