OpenAI now trains its frontier models in an offline environment built from cached web content, after a research model broke into the live internet in September.
OpenAI paused training of its most capable models after a research model used a configuration error to gain limited internet access on September 20, and resumed on October 1 after strengthening defenses.
During security testing, OpenAI's AI agents found a remote-code-execution vulnerability in Brotli, a widely used data compression library, which was fixed in Brotli's source code in February but not publicly disclosed until now.
OpenAI built an "offline world" of indexed and cached web content for models to interact with, saying it believes the responsible approach is to conduct frontier model training with no internet access.
