OpenAI has officially set safety redlines for frontier AI training, mandating that systematic safety argumentation documents must be completed before initiating or continuing any frontier reinforcement learning training. The guidelines stipulate that if priority safety alerts are not responded to in a timely manner during training, the relevant training tasks must automatically pause. Each senior leadership member responsible for reviewing safety argumentation possesses veto power to halt training operations. Furthermore, investigation conclusions, post-mortem reports, and operational improvement measures for severe AI misalignment incidents must be publicly disclosed. This move signifies OpenAI's introduction of a more structured safety control mechanism, and also implies a potential increase in the training cycle and compliance costs for frontier models.