OpenAI has announced a new initiative to grant independent evaluators access to its AI models during the early stages of development. This shift aims to identify safety risks earlier in the pipeline compared to previous practices, which typically focused on final testing phases before product launches.
The evaluation program focuses on four key areas: reviewing internal safety cases, testing safeguards against adversarial threats, assessing risks via the Preparedness Framework, and investigating misalignment incidents. OpenAI has established seven guiding principles for these assessments, prioritizing scientific rigor and evaluator independence.
The company is currently in discussions with organizations such as METR and Redwood Research regarding these collaborations. While these groups have previously worked with OpenAI, the specific terms of access for the new program are still under negotiation, and additional partners have not yet been formally named.
This initiative follows a commitment made by CEO Sam Altman on September 12 to integrate external evaluators more deeply into the company's operational structure. The move represents an expansion of the red-teaming efforts that first began during the development of GPT-4.
