OpenAI announced that all frontier reinforcement learning model training runs will occur within an offline environment isolated from the public internet