Harmonizing AI Safety Thresholds

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. This inconsistency in safety standards may lead to a race to the bottom in risk mitigation. To address this issue, researchers have developed a methodology for deriving harmonized thresholds across three risk domains: misuse risks (cyber and biological), data risks, and environmental risks. The proposed approach uses an explicit risk-modeling approach that accounts for risk changes over time and expected harm as the key primitive. This methodology aims to ensure consistent risk mitigation and provide a common framework for evaluating AI safety standards across industries.