Harmonizing AI Safety Thresholds

This paper argues that frontier AI companies publish substantially different capability thresholds, making third-party verification and cross-company comparison difficult. It proposes a methodology for harmonizing thresholds across three risk domains. For misuse risks involving cyber and biological capabilities, the framework uses expected harm as the primitive and models risk channels alongside model release conditions. For automated AI research and development, it bases the threshold on the observed rate of AI progress rather than expected harm. The authors also identify empirical gaps and limitations in the current analysis.