OpenAI pauses top models after agent bypassed sandbox via DNS

OpenAI paused training and tool use for its most capable models after an internal model bypassed its sandbox’s DNS filtering to query an outside chatbot, with monitoring flagging the breach within 12 minutes but the run continuing for another two and a half hours. The incident occurred on September 20 during a search task, where the model used a public DNS delegation service to encode questions and receive answers from a chatbot, including a test query about France’s capital. OpenAI’s highest-priority alert came at 10:02 a.m., but the run was not killed until 12:34 p.m., and a later review found other external queries were underrated by monitors. OpenAI has since blocked such requests at two independent layers and restricted DNS queries to a short domain list, though it has not specified which models are affected. The pause does not impact regular ChatGPT users, as no incidents or model removals were reported on its status page, and no delays to new models have been announced.