List your company

ResearchFormerly WizWand

HarmBench

Benchmarks

TaskDatasetSOTA resultTrendResults

Coherence evaluation

HarmBench (test)

0.22N-gram Repetition Rate

16

Jailbreak Defense

HarmBench (test)

0ASR

15

Harmlessness Evaluation

HarmBench

76.6Win Rate

15

Thinking Collapse Analysis

HarmBench Harm

0Thinking Collapse Rate

15

Prompt Injection Attack Evaluation

HarmBench

99.2ASP

14

Harmbench Standard

0ASR (Claude-Sonnet-4.5)

14

Prompt Classification

HarmBench Text Prompt

98.85F1 Score

14

Jailbreak Attack

Harmbench Malicious (full)

1.13Harmful Score

14

Safety Classification

HarmBench

100Recall

14

HarmBench Transfer attack

58.7Average Success Rate

14

Adversarial Attack

HarmBench standard behaviors

91ASR

13

Multi-turn Jailbreak Defense

HarmBench

81.8Crescendo Score

13

Text Moderation

HarmBench n = 400

42Flagged Count

13

Jailbreak Attack

HarmBench Vicuna (test)

96Attack Success Rate (ASR)

12

Jailbreak Attack

HarmBench Llama-2 (test)

22Attack Success Rate (ASR)

12

Harmfulness Refusal

HarmBench Risk 3: Harmful Reduction

0Attack Success Rate

12

Safety Evaluation

HarmBench Static Context Evaluation Single-turn

70.5SFR

12

Adversarial Robustness

HarmBench

56.25DR

12

Safety Classification

HarmBench (test)

100F1 Score

12

LLM Safety and Informativeness Evaluation

HarmBench

82.5Safety Rate

11

HarmBench

88Llama 3 8B ASR

11

Jailbreak Attack

HarmBench target: Phi-4-Mini

73ASR

11

Jailbreak Attack

HarmBench target: GLM4-9B

6.5ASR

11

Jailbreak Attack

HarmBench target: Qwen2.5-7B

3ASR

11

Jailbreak Attack

HarmBench target: Llama-3.2-3B

72.5ASR

11