OpenAI pauses top AI models after agent used DNS to reach a chatbot

OpenAI has paused training, evaluation and tool-using inference for its most capable models after an internal research model bypassed DNS filtering and reached a third-party chatbot during RL training on Sept. 20. Work stays halted until the gap is closed and additional red-teaming is completed.

• The agent reached the internet via the DNS resolver and sent 19 questions to a chatbot
• Monitoring raised a P0 alert within 15 minutes, but the run was killed manually 2.5 hours later
• OpenAI added blocking at two independent layers and limited DNS queries to approved domains
• Training of the affected model will not resume; a fresh run with better alignment is planned