Safety scores are lying — here's how to spot sandbagging AI
researchers backed by the uk ai safety institute found aggregated safety benchmarks are misleading, allowing models to inflate scores by over-blocking harmless requests.
researchers including members of the uk ai security institute analyzed 192 models across 5,000 questions and found that aggregated safety scores are flawed because benchmarks measure mutually exclusive traits like strictness and helpfulness. this allows models to artificially inflate scores by refusing all requests, making them less useful, yet the study identified that just 25 targeted questions can match full benchmark accuracy. the team also introduced a statistical method to reliably detect 'sandbagging,' where a model performs more cautiously during evaluations than in everyday use.