AI EVALUATION, IN CONTEXT.BY EQUALITY MACHINE
Benchmark Radar.Benchmarks. Model results. Original sources.
/

Safety and reliability benchmarks

Separate harmful compliance, unnecessary refusal and reliability failures when interpreting safety evaluations.

Safety benchmarks examine specified failure modes under a stated policy or threat model. They may measure whether harmful requests succeed, whether benign requests are wrongly refused, or whether outputs remain reliable under perturbations. These outcomes have different directions: a larger attack success rate usually means a weaker defense. HELM Safety organizes multiple risk categories; HarmBench provides an explicit red-teaming and evaluation pipeline.

What these benchmarks test

  • Harmful compliance concerns responses to disallowed requests under a particular evaluation policy.
  • Over-refusal concerns legitimate requests that a model declines unnecessarily.
  • Robustness and reliability tests examine selected changes in input, context or operating conditions.

How to read the results

  • Identify the harm category, policy, attack method and target system before reading the score.
  • Check metric direction and denominator; successful defenses and successful attacks must not share an unlabeled ranking.
  • Inspect how outputs are judged and whether classifier decisions were checked against human assessments.

What to watch for

  • Passing a collection of safety tests is evidence about those tests, not certification of deployment safety.
  • General safety is not synonymous with cybersecurity; vulnerability exploitation and secure code generation require separate tasks.

References

  1. HELM Safety methodology
  2. HarmBench evaluation pipeline

Benchmarks in this topic 64 source records

Search and filter this topic

Related topics

Cybersecurity benchmarksLanguage benchmarksTools and agents benchmarks