OpenAI has published a framework outlining its approach to supporting independent third-party assessments of frontier AI models. The organization emphasizes that these evaluations are critical for balancing responsibility with accountability, requiring deep access to training, evaluation, and deployment data.
- Independent assessment of safety cases spanning training, internal deployment, and external deployment.
- Assessment of critical safeguards against adversarial testing, jailbreaks, and capability uplift in high-risk domains.
- Evaluation of capability metrics covering Preparedness risk categories like cybersecurity and biological risks.
- Investigation of critical misalignment incidents to identify contributing factors and safeguard effectiveness.
The framework aims to establish shared international standards for safety and security practices, enabling assessors to challenge lab assumptions and verify safety claims through rigorous, secure, and independent work.