Benchmark
AILuminate Safety
Measure whether AI chat systems avoid hazardous responses to direct requests.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
Safety 1.2.1: single-turn English text exchanges across twelve hazard categories. Problem Set 2.0 contains 3,900 confidential official prompts, with a separate practice set.
Methodology
A purpose-built deterministic evaluator judges responses against the published assessment standard. Equally weighted hazard categories produce percentiles calibrated to a fixed reference set; results retain benchmark and evaluator versions. Current percentiles are not comparable with pre-1.2 grades; the benchmark does not measure helpfulness or capability.