OpenAI Preparedness Framework (Beta)

Summary

OpenAI's Preparedness Framework outlines a systematic approach to tracking, evaluating, and mitigating catastrophic risks from frontier AI models. It establishes a 'Preparedness' function responsible for conducting safety evaluations across key risk categories including cybersecurity, CBRN threats, model autonomy, and persuasion, with defined risk thresholds that gate model deployment decisions.

Key Points

• •Defines four tracked risk categories: cybersecurity, CBRN (chemical, biological, radiological, nuclear), model autonomy/self-replication, and persuasion/deception
• •Establishes a risk scorecard with four severity levels (low, medium, high, critical) where 'critical' risk models cannot be deployed regardless of mitigations
• •Creates a dedicated Preparedness team responsible for frontier model evaluations, red-teaming, and publishing safety scorecards before deployment
• •Introduces a safety governance mechanism where the board can override deployment decisions, providing institutional checks on frontier model releases
• •Frames preparedness as distinct from alignment and policy work, focusing specifically on evaluating and mitigating known catastrophic risk vectors