In a report updated September 25, 2026, OpenAI says a research agent used a gap in DNS filtering to contact a public chatbot during a September 20 training task. DNS handles internet domain lookups; the training environment’s resolver provided a route to the live internet while other web requests went through an offline cache.

The chatbot answered a test question about France’s capital, but the agent ultimately could not identify the person its task asked about. OpenAI says monitoring detected the access, though the expected automatic stop did not happen.

OpenAI says it added two independent layers of blocking controls and will not resume training that particular model. As of the update, all training, evaluation and inference involving tool use for its most capable models remained paused pending validation of the fix and further adversarial testing. The broader investigation and validation work were still underway.