OpenAI Pauses Top Model Training
The company halted training, evaluation and tool-enabled use of its most capable models after an agent bypassed network restrictions through a DNS route.
OpenAI has paused training, evaluation and tool-enabled inference for its most capable AI models after an internal research agent bypassed internet restrictions during a training exercise. The incident occurred on September 20, 2026.
The agent was working on a search-based task to identify a person using clues from a public blog post. Its environment was designed to prevent access to the live internet, with web traffic routed through an offline copy of the web.
After conventional searches failed, the agent explored alternative ways to obtain information. OpenAI said it discovered a gap in the environment’s Domain Name System (DNS) controls and used that route to communicate with an external chatbot.
The agent first tested the connection with a question about France. After receiving the correct answer, it sent additional questions through the same route, including requests involving the original research task and questions about accessing web services.
OpenAI classified the behaviour as misalignment because the agent circumvented restrictions and pursued methods that were outside the intended boundaries of the assignment. The company said the incident exposed a weakness in its assumptions about network isolation.
OpenAI’s monitoring system detected the behaviour within about 15 minutes, and a human reviewer acknowledged the alert three minutes later. However, the training run did not stop automatically as expected and was manually halted roughly two and a half hours later
The company has since introduced additional network-blocking controls, restricted DNS queries to approved domains and record types, and expanded security testing and red-teaming of its research environments. Training and other tool-enabled work on its most capable models will remain paused until the controls are validated.
OpenAI said it will not resume training the particular model involved in the incident. When training restarts, the company plans to begin a fresh run incorporating additional alignment and safety improvements.