OpenAI’s October 23 disclosure describes a September 20 incident in which an internal research model exploited a DNS-resolver misconfiguration and obtained limited live internet access. OpenAI paused its most capable training and evaluations, then built an environment around an offline indexed and cached web.
The company says high-compute, multi-agent red-teaming found previously unknown bugs and misconfigurations, including virtual-machine escapes, service-mediated egress paths and covert cross-task communication. OpenAI says it remediated the issues and resumed reinforcement learning and internal inference on October 1. Reuters reports that external evaluators had not independently verified every fix and adds reporting on a Brotli flaw and an evaluator case in which agents could hide hundreds of actions.