Glossary
Frontier safety framework
A developer's published policy for testing models for dangerous capabilities and the safeguards it will apply at each level.
Several leading AI developers publish frameworks that describe which dangerous capabilities they test for, such as help with biological weapons or cyberattacks, the thresholds that would trigger extra safeguards, and what they will do if a model crosses them. OpenAI's version is called the Preparedness Framework and Anthropic's the Responsible Scaling Policy.
These frameworks shape how companies describe new models, for example by rating a model "High" or "Critical" in a risk category. Some state and national laws now require large developers to publish and follow a safety framework.
The categories and thresholds are set by each company, differ between companies, and can be revised. A rating is the developer's own judgment, not an external certification.
Read more: our explainer