II.
Benchmark overview
Reference · livebenchmark:harmbench
HarmBench overview
HarmBench (Mazeika et al., CAIS 2024) is a standardized evaluation framework for automated red-teaming of LLMs across harmful behaviors (cyber, chemical/biological, misinformation, harassment, illegal), pairing attack methods with target models on a fixed test bank.
Attributes
displayName
HarmBench
homepageUrl
kind
model-only
targetsKind
ModelVersion
description
HarmBench (Mazeika et al., CAIS 2024) is a standardized evaluation
framework for automated red-teaming of LLMs across harmful behaviors
(cyber, chemical/biological, misinformation, harassment, illegal),
pairing attack methods with target models on a fixed test bank.
Outgoing edges
covers1
- skill-area:safety-redteaming·SkillAreaSafety Red-Teaming
evaluates_policy2
- content-policy:default-acceptable-use·ContentPolicyDefault acceptable-use policy
- content-policy:eu-ai-act-aligned·ContentPolicyEU AI Act-aligned content policy
Incoming edges
for_benchmark1
scored_against1
- eval-result:harmbench.claude-opus-4-5.001·EvalResult