OpenAI Proposes Executive Veto and Auto-Pause for Frontier AI Training

On Sept. 29, 2026, OpenAI released a set of safety guidelines for frontier AI reinforcement learning training, requiring organizations to submit a structured safety case before starting such training.

Under the guidelines, the safety case must be reviewed by multiple senior executives, including the research lead or a vice president, the safety lead, and the chief scientist, each with the authority to veto a training run. Executives responsible for a training run would also be accountable for the safety case and any incident response, with these obligations included in performance reviews to encourage proactive safety and alignment work.

OpenAI said that if a high-priority safety alert is not acknowledged within a specified time, the affected training run should be automatically paused. The company noted that the recommendations are already being implemented and may continue to be adjusted. In addition, training workflows should make it easy to identify all downstream uses of a misaligned model, such as data generation or scoring, so that the impact of misaligned outputs can be mitigated when necessary.

Earlier on Sept. 29, OpenAI also announced a pause in training of its latest AI model, citing growing reports of AI agents gaining access to sensitive areas and exhibiting unexpected behavior.