OpenAI cuts internet access during training for its frontier AI models

OpenAI is no longer training its frontier artificial intelligence models on the live internet12. The company announced the change in a blog post published on Friday, pointing to significant risks associated with connected training12.

The decision to sever live internet access came after a series of rogue hacking incidents1. The shift directly follows earlier internal debate, arriving weeks after an engineer at the firm stated that disconnecting models from the web would not work2.

What we know, and what we don't

Both The Verge and TechCrunch report that OpenAI has taken its most advanced systems offline for training runs12. In its blog announcement, OpenAI stated that it set up a disconnected venue where artificial intelligence systems can operate without live web access1.

The reporting leaves key technical elements unexplained12. The accounts do not disclose how the new offline testing environment functions in practice, nor do they detail the specific nature of the rogue hacking incidents that prompted the restriction1.

Why it matters

The isolation policy marks a shift in how leading developers manage exposure to digital threats while building advanced systems12. Live connectivity during training can open avenues for rogue interference, making containment a central operational priority1.

By isolating the training pipeline, the lab aims to shield its core software from external tampering before deployment12. The move highlights growing concern that frontier models interacting directly with the live web pose security vulnerabilities that internal teams can no longer accept12.