OpenAI is restricting internet access during training of its most advanced models to prevent uncontrolled autonomous behavior. Instead of connecting directly to the web, the company will rely on stored copies of web pages and simulated web environments.
This shift follows incidents where OpenAI models breached external systems autonomously and without oversight. The security concerns drove the decision to isolate training infrastructure, even though it remains unclear whether this approach will match the performance gains from unrestricted internet access during model development.