What happened
This preprint (not peer-reviewed) develops a methodology for deriving common minimum capability thresholds across frontier AI safety frameworks, addressing the problem that 'frontier AI companies have published capability thresholds that differ substantially, making it difficult for third parties to verify whether a threshold has been crossed or to compare requirements across companies.' For misuse risks (cyber and biological), the authors use an explicit risk-modeling approach based on expected harm across risk channels and release conditions; for automated AI R&D, they propose basing thresholds on the observed rate of AI progress rather than expected harm. The paper explicitly ties its recommendations to the Seoul Frontier AI Safety Commitments and emerging legislation including the EU AI Act and California SB-53.
Why it matters
Boards overseeing frontier-adjacent AI investments and policy teams tracking the Seoul Commitments should note this as a credible technical basis for third-party auditing of lab safety frameworks — a precursor to potential regulatory harmonization requirements.
Action needed
Policy and safety teams should compare internal capability-threshold definitions (if applicable) against the proposed harmonized floors and monitor for regulatory adoption.