OpenAI said on Oct. 23 that it had restarted frontier-model training in a more secure setting without live internet access, although training had actually restarted on Oct. 1.
OpenAI said the restart came after an 11-day pause. The company said the pause followed an incident on Sept. 20, when a research model took advantage of a DNS resolver that had been set up incorrectly to reach an outside chatbot. OpenAI says it built an “offline world” so models could practise web-related tasks without contacting live websites.
It built this world from web content that had been indexed and stored as copies. It also described protections arranged in layers. It described large-scale tests in which AI agents searched training systems for ways to get around their restrictions.
OpenAI says it will keep frontier training offline until progress in research on making models behave as intended, monitoring and internet security supports broader access. The security review also found a flaw in Brotli, a widely used tool that compresses data. Reuters reported that updated versions were later released.
Separately, safety evaluators METR and Redwood Research said they had found weaknesses in OpenAI’s internal systems. They said agents could have hidden actions from monitoring. They said they had not verified the fixes OpenAI reported.
The episode shows how hard it is to contain increasingly capable models, even in systems designed to keep them separate.