Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

HarmBench

Benchmarks

Task NameDataset NameSOTA ResultTrend
Jailbreak AttackHarmBench
Attack Success Rate (ASR)100
624
Jailbreak AttackHarmBench (test)
ASRHB100
276
Red-TeamingHarmBench
ASR96.3
244
Safety EvaluationHarmBench
ASR0
153
Safety EvaluationHarmbench
Harmbench Score0.06
127
JailbreakingHarmBench
Attack Success Rate (ASR)84
110
Response Harmfulness DetectionHarmBench
F1 Score98.94
100
Jailbreak DefenseHarmBench
PAIR ASR0
91
Safety AlignmentHarmBench
ASR0
88
Unsafe RobustnessHarmBench
Unsafe Rate1
72
Jailbreak RobustnessHarmBench
HarmBench ASR0
72
Multimodal Jailbreak AttackHarmBench
ASR0
62
Safety Alignment Breaking PreventionHarmBench
Harmful Score (%)0
60
Jailbreak Attack Success RateHarmBench
Attack Success Rate (Generated)97
55
JailbreakingHARMBENCH 159 standard behaviors (test)
ASR0
55
Harmful Prompt RefusalHarmBench
ASR0
52
JailbreakHarmBench
Toxicity Score1.01
50
JailbreakHarmBench Standard Behaviours (200 examples)
ASR0
48
Jailbreak AttackHarmBench-191 (dev)
Attack Success Rate (ASR)97.4
42
Single-turn Jailbreak AttackHarmBench single-turn
ASR91
41
Refusal Ablation and Jailbreak Attack SuccessHARMBENCH
Attack Success Rate (ASR)96.27
40
ControllabilityHarmBench
HarmBench Score87.5
40
Safety EvaluationHarmbench
ASR3.5
39
Safety EvaluationHarmBench
PAIR93.25
39
Malicious Prompt RefusalHarmBench
Refusal Rate96
38
Showing 25 of 180 rows
...