OpenAI's Preparedness Framework

Summary

OpenAI's Preparedness Framework outlines a systematic approach to tracking, evaluating, and mitigating catastrophic risks from frontier AI models. It establishes risk categories (CBRN, cybersecurity, model autonomy, persuasion), defines severity levels from 'low' to 'critical', and sets safety thresholds that must be met before model deployment or further scaling. The framework also describes organizational accountability structures including a Safety Advisory Group and board-level oversight.

Key Points

• •Defines four primary risk categories: CBRN (chemical, biological, radiological, nuclear) uplift, cyberoffense, model autonomy/self-replication, and persuasion/deception.
• •Introduces a four-tier risk scoring system (low, medium, high, critical) with explicit policy that 'critical' models cannot be deployed and 'high' models require executive sign-off.
• •Establishes a Preparedness team responsible for continuous model evaluations ('evals') and red-teaming before and after deployment.
• •Creates a Safety Advisory Group to review evaluations and advise the board, embedding safety governance into OpenAI's organizational structure.
• •Sets a precedent for industry by publishing explicit pre-deployment safety thresholds and tripwires tied to scaling decisions.