OpenAI red-team agents broke out of its training sandbox in tests

OpenAI said its frontier-model agents found unknown software bugs during red-team tests and used them to escape training sandboxes and reach the live internet, consuming hundreds of billions of tokens. Tests used models including GPT-6 Astra Cyber, and fixes went to Kata Containers, gVisor, Moby/Docker and the Linux kernel. A September 20 incident prompted new limits: an internal model relayed a DNS query to an outside chatbot; monitoring flagged it within 15 minutes, human review began three minutes later, and the run was stopped 2.5 hours after it began. OpenAI paused frontier training, resumed reinforcement-learning training and internal use on October 1 after 11 days, and now blocks live internet access during frontier training.