OpenAI published initial safety-case guidelines for frontier reinforcement-learning training on September 28. The proposals combine evidence about alignment, containment and monitoring with operational checks, including leadership review, audit access and procedures for pausing problematic runs.
The document also recommends investigating serious misalignment incidents and sharing findings after investigations conclude. OpenAI describes rigorous safety cases as a goal and says these practices are being implemented and will evolve. The scope is training; deployment requires broader assessment. Official guidelines.