OpenAI's risk evaluation framework
PolicyIntroduced OpenAI's framework for evaluating catastrophic risks from frontier models across four categories (cybersecurity, CBRN, persuasion, model autonomy), with risk levels that determine deployment decisions.
Low, Medium, High, Critical. Models at "High" can be deployed with mitigations. Models at "Critical" cannot be deployed.
The Safety Advisory Group reviews assessments. The board can overrule deployment decisions.