The Role of AI Safety Benchmarks in Evaluating Systemic Risks in General-Purpose AI Models

Abstract

The evaluation of systemic risks in General-Purpose AI (GPAI) models is a complex challenge that requires a multifaceted approach, extending beyond traditional capability assessments. This report analyses the role of AI safety benchmarks in identifying systemic risks in GPAI models, proposing a dual-trigger framework that combines capability triggers with safety benchmarks to provide a more comprehensive assessment of potential harms. The current landscape of safety benchmarks is still in development, with various initiatives emerging to address specific systemic risk categories, such as the ones identified in the GPAI Code of Practice: cyber offence, chemical, biological, radiological and nuclear (CBRN) risks, harmful manipulation, and loss of control. A tiered evaluation strategy is recommended, applying more rigorous and costly safety evaluations only to models that meet a predefined capability threshold or are intended for deployment in high-risk domains, ensuring proportionality and efficient resource allocation.