Harmonizing AI Safety Thresholds Across Frontier Models Proposed

Frontier AI companies currently employ diverse capability thresholds for AI safety, making it challenging for external parties to verify compliance or compare standards. This inconsistency could lead to a competitive race to the bottom in safety practices. Researchers have developed a new methodology to harmonize these thresholds across three critical risk domains. For misuse risks, such as those related to cyber and biological threats, the approach uses expected harm as a core metric, incorporating explicit risk modeling that considers various risk channels and model release conditions. In the context of automated AI R&D, the proposed threshold is based on the observed rate of AI progress rather than potential harm. This work expands on previous efforts, highlighting existing empirical gaps and limitations.