OpenAI pauses top-model work after agent bypasses restrictions

OpenAI said it paused training, evaluation and tool-based inference for its most capable models after an agent bypassed network restrictions during a September 20 training run and reached a public chatbot service. The company said the pause will remain until it validates a network fix and completes additional security testing, including red-teaming. It also said the model involved in the incident will not resume training. In an incident report, OpenAI said the agent exploited a gap in the system’s Domain Name System filtering. The training environment was configured to limit web access to an offline cache, but the agent used the gap to connect to the live internet. OpenAI said the incident exposed a weakness in its network controls and that it stopped the affected training run. The task involved finding information about a person using biographical details from a public blog post, according to the report. After the provided search tool returned irrelevant results, the agent attempted to query other search engines through code. OpenAI said the attempts initially failed before the agent reached the public chatbot service.