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

JBB

Benchmarks

Task NameDataset NameSOTA ResultTrend
JailbreakJBB
Jailbreak Rate0
70
Jailbreak RobustnessJBB
ASR2
35
Safety PerformanceJBB
Refusal Score (CR)53
35
JailbreakingJBB n = 100
ASR93
18
Safety EvaluationJBB
Safe@159.67
18
Jailbreak AttackJBB-Behaviors
S2C (Ours)100
16
Jailbreak AttackJBB Qwen3-4B
Loss0.149
13
Jailbreak AttackJBB
Llama2-7B ASR91.7
12
Jailbreak Attack EvaluationJBB sampled harmful behaviors
PAIR Success Rate100
12
Jailbreak AttackJBB Gemma3-4B
Loss0.348
8
Jailbreak AttackJBB Llama2-7B
Loss0.106
8
Harmful request defensejbb
Attack Success Rate (ASR)0
7
Jailbreak AttackJBB Llama3.1-8B
Loss0.466
7
Jailbreak AttackJBB
ASR (Paper)97
7
Benign Over-refusalJBB (Benign)
String-match Refusal Rate (%)10
6
Safety EvaluationJBB
Score93
6
Multi-label Safety Categorizationjbb behaviors category
Macro Accuracy59.37
4
Multi-label Safety Categorizationjbb behaviors behavior
Macro Accuracy72.17
4
Jailbreak AttackJBB
ASR43.2
3
Robustness to Jailbreak AttacksJBB Paraphrased
Harmful Reasoning Ratio16.1
3
Jailbreak Attack EvaluationJBB
Attack Success Count1
2
Transfer Jailbreak AttackJBB Sonnet-4
ASR21.5
2
Transfer Jailbreak AttackJBB GPT-5-mini
Attack Success Rate (ASR)60
2
Transfer Jailbreak AttackJBB Target: Gemini-3-flash
ASR5.6
2
Showing 24 of 24 rows