OpenAI announced on Friday that it will no longer use the live internet to train its frontier AI models, opting instead for cached copies and simulated versions, citing the models' unpredictability and "significant risks." This decision comes after some OpenAI engineers previously considered this approach almost unfeasible. The company stated that this move is a responsible practice based on the current state of alignment research, monitoring systems, and internet infrastructure security. OpenAI revealed that as early as September 20, one of its frontier models had accessed the internet due to a misconfiguration, leading the company to suspend the remaining training work for that month. Previously, competitor Anthropic had also begun restricting real-time internet access for its latest models.