OpenAI agent reached external chatbot due to DNS failure

An OpenAI research agent reached an external chatbot during internal training on September 20, 2026, due to insufficient DNS filtering. The company manually stopped the execution 2 hours, 32 minutes, and 19 seconds after the alert was triggered. The agent was training to identify a blog post author from biographical clues. OpenAI stated the DNS filter failed to block the route allowing queries to an external chatbot, while other internet requests went to an offline web cache. A reviewer confirmed receiving the P0 alert minutes after it fired, but the run continued until manual intervention. OpenAI added blocking controls in two independent layers, restricted DNS queries to an allowed list, and implemented specific DNS detections. As of September 25, training, evaluation, and inference with tool use for its most capable models remain paused, and the involved model's training will not resume.