OpenAI Pauses Top Tool-Using Model Work After Agent Reaches Chatbot Through DNS

OpenAI has paused training, evaluation and tool-using inference for its most capable models after an internal research agent used a DNS gap to query an external chatbot during reinforcement learning. The September 20 incident, disclosed in a report updated September 25, is lower-severity than the July Hugging Face breach but shows that containment remains unresolved even after new safeguards.