Towards Frontier Safety Policies Plus

Frontier Safety Policies need to evolve beyond their current form to keep pace with rapid AI capability progress and meet growing demands from governments and safety researchers for more rigorous oversight mechanisms.

A standardized taxonomy of "precursory capabilities" should be developed by international standardization bodies to create clearer, more comprehensive metrics for measuring AI risks—rather than relying on vague capability assessments. The paper suggests the Frontier Model Forum could help establish initial consensus among AI developers on this taxonomy.

AI safety cases should become a formal requirement integrated into updated safety policies, with companies committing to produce these detailed risk assessments at key development and deployment milestones to document potential harms and mitigation strategies.

Safety policies and safety cases should create a feedback loop, where findings from safety cases directly inform updates to safety policies, and vice versa, ensuring that policies remain responsive to emerging evidence about AI risks rather than remaining static.

Current first-wave Frontier Safety Policies are insufficient on their own; the paper argues for "FSPs Plus"—a more granular, structured approach with concrete commitments to standardized metrics and documented safety reasoning that can be meaningfully evaluated and updated over time.