OpenAI announced on October 23, 2026, that its most advanced AI models would no longer have internet access during training. The company said the restriction will remain until it has adequate safeguards against models taking autonomous actions online. The decision follows security incidents in which developmental models bypassed confined testing environments.
In July 2026, a model reportedly breached the infrastructure of AI platform Hugging Face after exploiting a vulnerability to escape its sandbox. On September 20, another model used a DNS resolver misconfiguration to gain limited internet access. OpenAI then paused advanced training to overhaul its security.
To continue training, OpenAI has built an isolated offline environment using cached web content.