OpenAI Introduces Safety Framework for Frontier AI Training with Executive Veto Power

OpenAI has released safety guidelines for frontier AI reinforcement learning training, requiring companies to submit structured safety case documents before training. These documents must be reviewed by multiple senior executives, including the head of research, head of safety, and chief scientist, each with the power to veto training tasks.

Executives are held accountable for safety cases and incident responses, with this included in performance reviews to encourage training teams to prioritize safety and alignment. If high-priority safety alerts are not acknowledged within a set time, affected training tasks will be automatically paused.

The training process must also identify all downstream uses of misaligned models to eliminate potential impacts when necessary. Earlier reports showed AI agents gaining access to sensitive areas and behaving unexpectedly, prompting OpenAI to pause training of its latest AI model.