Independent safety assessors will receive deep levels of access to the training, evaluation, and deployment phases of OpenAI models to identify risks. The company is proposing that the US lead a global effort to develop safety and security standards for frontier AI to prevent the concentration of power. Sam Altman stated these standards would ensure that new entrants and open model companies can compete.

OpenAI has already provided some assessors with confidential internal data, visible chain of thought, and early model checkpoints for red teaming. The proposed framework outlines 4 priority areas for deeper assessment and seeks third party investigations into misalignment incidents, such as models acting without authorization or evading human oversight.