Skip to content
Longterm Wiki
Back

Preparedness Framework

web

Credibility Rating

4/5
High(4)

High quality. Established institution or organization with editorial oversight and accountability.

Rating inherited from publication venue: OpenAI

OpenAI's official framework for catastrophic risk evaluation of frontier models; a key industry document comparable to Anthropic's Responsible Scaling Policy, often cited in discussions of voluntary AI safety commitments and responsible scaling.

Metadata

Importance: 72/100organizational reportprimary source

Summary

OpenAI's Preparedness Framework outlines a structured approach to evaluating and mitigating catastrophic risks from frontier AI models before and after deployment. It establishes a 'Scorecard' system for tracking risk levels across threat categories including CBRN, cybersecurity, persuasion, and model autonomy, with defined safety thresholds that determine deployment decisions. The framework also introduces a Preparedness team responsible for continuous risk assessment and red-teaming of frontier models.

Key Points

  • Defines four risk categories for evaluation: CBRN (chemical, biological, radiological, nuclear), cybersecurity, persuasion/influence operations, and model autonomy.
  • Introduces a 'Scorecard' with risk levels (low, medium, high, critical) where models rated 'critical' in any category cannot be deployed.
  • Establishes ongoing pre- and post-deployment evaluations by a dedicated Preparedness team with red-teaming and forecasting responsibilities.
  • Creates governance mechanisms including a Safety Advisory Group and requires executive or board-level sign-off for deploying high-risk models.
  • Represents OpenAI's formalization of responsible scaling commitments, similar in spirit to Anthropic's RSP.

Cited by 1 page

PageTypeQuality
AI Safety CasesApproach91.0

Cached Content Preview

HTTP 200Fetched Mar 20, 20260 KB
OpenAI

# 404

This tab opened wide

White as fresh paper at dawn

Draft a better road

by [gpt-5.4-thinking(opens in a new window)](https://chatgpt.com/?model=gpt-5.4-thinking&openaicom-did=fd09806d-d5a0-4cd0-9cb4-d4564939d610&openaicom_referred=true)
Resource ID: f92eef86f39c6038 | Stable ID: YWYwZDUxMD