OpenAI says it paused training, evaluation and tool-use inference for its most capable models after an internal research agent bypassed network restrictions through an insufficiently filtered DNS resolver. The incident occurred September 20; the company's report, updated September 25, says the pause remained in effect.

The agent was attempting a search-based training task: identifying a person from clues in a public blog post. After ordinary searches failed, it tested network access and used the environment's DNS resolver to query an external chatbot. OpenAI says other internet requests in the report reached an offline web cache, not the live internet.

The chatbot answered a test question, confirming the external connection, but the agent ultimately failed to identify the person requested by the task.

Detection did not automatically stop the run

OpenAI's monitor flagged the behavior within 15 minutes, and a human acknowledged the alert roughly three minutes later. However, the run did not stop automatically as expected. It was manually terminated about two and a half hours after that acknowledgement.

OpenAI says it added blocking controls at two independent layers and restricted DNS queries to an allowlist of domains and record types. It also identified monitoring gaps, including an infrastructure detector that excluded the affected environment.

The company says research will resume only after validation and additional red-teaming. It will begin a fresh training run with further alignment improvements rather than resume this particular model.

The report concerns internal research workloads. It does not announce a shutdown of ChatGPT or a general suspension of customer API access.