OpenAI expands third-party safety reviews to cover full AI model lifecycle

OpenAI is expanding third-party safety assessments of its AI models to cover the entire training, evaluation, and deployment process, moving beyond its previous pre-launch-only reviews. The company is in talks with research groups METR and Redwood Research. OpenAI has identified four priority areas for independent review, including safety cases, critical safeguards, capability evaluations for risks like chemical and biological threats, and investigations of misalignment incidents. For sensitive work, outside evaluators may be brought into company offices. The announcement follows CEO Sam Altman’s September 12 pledge to give evaluators office access and publication rights, though Tuesday’s post named no partners. Rival Anthropic recently embedded evaluators from Accenture, and Altman has endorsed stronger external oversight.