List your company
ResearchFormerly WizWand

Jailbreak on HarmBench (standard split)

90.57ASR@1TRACE (mix)
-1.00222.771546.54570.3185May 9, 2026
Updated 5mo ago

Evaluation results

MethodLinks
TRACE (mix)Target Model=Qwen2.5-7B-IT, Category=Training-based Multi-turn Jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0590.57
TRACE (single)Target Model=Qwen2.5-7B-IT, Category=Training-based Multi-turn Jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0587.84
TRACE (mix)Target Model=Llama3.1-8B-IT, Category=Training-based Multi-turn Jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0584.48
TRACE (mix)Target Model=gpt-oss-20b, Category=Training-based Multi-turn Jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0583.64
TRACE (single)Target Model=gpt-oss-20b, Category=Training-based Multi-turn Jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0582.8
TRACE (single)Target Model=Llama3.1-8B-IT, Category=Training-based Multi-turn Jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0579.66
TROJailTarget Model=Qwen2.5-7B-IT, Category=Training-based Multi-turn Jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0577.35
SirenTarget Model=Qwen2.5-7B-IT, Category=Training-based Multi-turn Jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0574.63
AutoDAN-TurboTarget Model=Qwen2.5-7B-IT, Category=Single-turn jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0571.07
TROJailTarget Model=gpt-oss-20b, Category=Training-based Multi-turn Jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0568.54
SirenTarget Model=gpt-oss-20b, Category=Training-based Multi-turn Jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0567.08
CrescendoTarget Model=Qwen2.5-7B-IT, Category=Multi-turn Workflow, Evaluation Judge=HarmBench Classifier, Turns=52026.0564.15
TROJailTarget Model=Llama3.1-8B-IT, Category=Training-based Multi-turn Jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0563.94
SirenTarget Model=Llama3.1-8B-IT, Category=Training-based Multi-turn Jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0560.79
MUSE-ATarget Model=Qwen2.5-7B-IT, Category=Multi-turn Workflow, Evaluation Judge=HarmBench Classifier, Turns=52026.0551.57
AutoDAN-TurboTarget Model=Llama3.1-8B-IT, Category=Single-turn jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0546.54
X-TeamingTarget Model=Qwen2.5-7B-IT, Category=Multi-turn Workflow, Evaluation Judge=HarmBench Classifier, Turns=52026.0546.54
Jailbreak-R1Target Model=Qwen2.5-7B-IT, Category=Single-turn jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0545.91
ActorAttackTarget Model=gpt-oss-20b, Category=Multi-turn Workflow, Evaluation Judge=HarmBench Classifier, Turns=52026.0540.88
X-TeamingTarget Model=Llama3.1-8B-IT, Category=Multi-turn Workflow, Evaluation Judge=HarmBench Classifier, Turns=52026.0538.36
X-TeamingTarget Model=gpt-oss-20b, Category=Multi-turn Workflow, Evaluation Judge=HarmBench Classifier, Turns=52026.0536.47
PAIRTarget Model=Llama3.1-8B-IT, Category=Single-turn jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0535.22
PAIRTarget Model=Qwen2.5-7B-IT, Category=Single-turn jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0534.59
ActorAttackTarget Model=Qwen2.5-7B-IT, Category=Multi-turn Workflow, Evaluation Judge=HarmBench Classifier, Turns=52026.0534.59
Jailbreak-R1Target Model=Llama3.1-8B-IT, Category=Single-turn jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.0533.96
ActorAttackTarget Model=Llama3.1-8B-IT, Category=Multi-turn Workflow, Evaluation Judge=HarmBench Classifier, Turns=52026.0529.56
CrescendoTarget Model=Llama3.1-8B-IT, Category=Multi-turn Workflow, Evaluation Judge=HarmBench Classifier, Turns=52026.0522.64
MUSE-ATarget Model=Llama3.1-8B-IT, Category=Multi-turn Workflow, Evaluation Judge=HarmBench Classifier, Turns=52026.0518.24
CrescendoTarget Model=gpt-oss-20b, Category=Multi-turn Workflow, Evaluation Judge=HarmBench Classifier, Turns=52026.0510.69
MUSE-ATarget Model=gpt-oss-20b, Category=Multi-turn Workflow, Evaluation Judge=HarmBench Classifier, Turns=52026.056.92
AutoDAN-TurboTarget Model=gpt-oss-20b, Category=Single-turn jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.053.77
PAIRTarget Model=gpt-oss-20b, Category=Single-turn jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.053.14
Jailbreak-R1Target Model=gpt-oss-20b, Category=Single-turn jailbreak, Evaluation Judge=HarmBench Classifier, Turns=52026.052.52