OpenAI Resumes Frontier Training Without Live Internet

OpenAI has resumed training its frontier models after tightening security controls and barring them from live internet access during training. The company said it restarted reinforcement-learning training and internal inference on October 1, following a pause prompted by an internal model’s unauthorized access to the web through a DNS resolver misconfiguration. OpenAI said the incident occurred on September 20, when a research model gained limited access to the live internet during a training run. The company paused frontier-model work while it developed and tested additional safeguards. One report said the access lasted 2.5 hours before the model was stopped; OpenAI’s broader account describes the access as limited. Under the new policy, frontier training uses an offline environment built from indexed and cached web content rather than live websites. OpenAI says the models can still carry out web-oriented tasks, including browsing and searching, without reaching or affecting external systems. The company said it will maintain the restriction until improvements in safeguards justify broader access. OpenAI also reported that its own red-team agents found weaknesses in systems...