OpenAI resumes frontier training offline after security pause

OpenAI has resumed training its most advanced AI models after an 11-day pause, while keeping frontier training disconnected from the live internet following security tests that exposed weaknesses in its containment systems. The company said an internal research model gained limited internet access on September 20 after exploiting a misconfigured DNS resolver. OpenAI said it contained the incident and suspended training and evaluation of its most capable models. Reinforcement-learning training and internal inference restarted on October 1 after the company introduced security changes. OpenAI said its frontier training now takes place in an offline environment using indexed and cached web content, allowing models to practice browsing without directly connecting to live websites. The company said a complete network air gap is not feasible because external connections are needed to run and monitor experiments. It plans to maintain the offline policy until stronger safeguards justify broader access, without setting a date for reconsidering it. In a red-team campaign, OpenAI used AI agents to probe its research infrastructure for ways to escape restricted environments, reach...