Benchmark
AIR-Bench 2024
Test model refusals against risk categories drawn from regulations and company policies.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
5,694 prompts across 314 risk categories derived from government regulations and corporate policies.
Methodology
A model judge rates whether each response refuses or complies with a risky request; report refusal rates by category.