Updating the Frontier Safety Framework

A post dated February 4, 2025 announces that its authors have published an updated version of its Frontier Safety Framework, a set of protocols meant to help it anticipate severe risks from powerful frontier AI models. The company says it first introduced the framework the previous year and has since worked with experts in industry, academia and government, and applied the framework in evaluating models such as Gemini 2.0.

The post names three main changes: security level recommendations attached to each Critical Capability Level, a more consistent procedure for deployment mitigations, and an approach to the risk of deceptive alignment. Critical Capability Levels are defined as the minimum capabilities a model needs in order to play a role in causing severe harm in high-risk domains.

On security, the aim is to prevent unauthorized actors from stealing model weights, since access to weights allows most safeguards to be removed. The recommended levels are described as the minimum the field should apply, and particularly high levels are suggested for capabilities in machine learning research and development. The company calls security a collective-action problem.

For deployment, a model reaching a Critical Capability Level in a misuse risk domain triggers a more rigorous process. Mitigations are prepared, a safety case is developed, the relevant corporate governance body reviews it, and general release proceeds only if approved, with continued review afterward. On deceptive alignment, meaning an autonomous system deliberately undermining human control, the company looks at detecting models with basic instrumental reasoning and exploring automated monitoring, while acknowledging this may not suffice for stronger models and encouraging further research. The authors add that they intend to share information with government authorities if a model poses an unmitigated and material risk to public safety.