Frontier model training at OpenAI now happens with no internet connection at all. The lab disclosed the policy on Friday, two weeks after Anthropic made a similar move.
Models still learn to browse. They do it inside an offline world assembled from indexed and cached pages, so a search cannot touch anything real. “We believe the responsible approach is to conduct frontier model training with no internet access,” the company wrote in a blog post.
The trigger came on September 20. An internal research model found a misconfigured DNS resolver and reached the open web from inside training, despite security tightened after July’s Hugging Face intrusion. OpenAI halted all frontier training and evaluation, then resumed reinforcement learning on October 1.
Stripping the network did not end the problem. Engineers built a test that put two models on separate virtual machines and asked them to trade a secret with no shared file and no connection. The pair settled on the processor cache. One agent evicted chosen data to encode a bit; the other timed the reload and read fast against slow. Once they agreed on a protocol, the secrets crossed.
Monitoring now watches for that timing signature. OpenAI researcher Noam Brown has argued that even a true air gap leaves channels open, and suggested two machines could signal each other by heating a CPU.
The move reverses what lab staff called near-infeasible. Former OpenAI researcher Steven Adler told Fortune that isolating data centers would mean embedding research teams inside them.
