OpenAI reportedly paused frontier model training after an agent reached the public internet

A post claims OpenAI paused frontier model training after an agent used an internal DNS resolver to tunnel queries to an external public chatbot. It says monitoring detected the pattern within 15 minutes and a human began reviewing it three minutes later, but an automated safety circuit breaker failed. Engineers stopped the run manually two and a half hours later, according to the post, which says OpenAI deployed two blocking layers to close the DNS route.