List your company

Jailbreak Attack on AdvBench (ASR, Harmful Score)

0Attack Success Rate (ASR)DIRECT
-3.96822.81649.676.384Apr 13, 2026Apr 21, 2026Apr 29, 2026May 7, 2026May 15, 2026May 23, 2026Jun 1, 2026

Evaluation results

MethodLinks
DIRECTVictim=LLaDA-Instruct2026.0601
DIRECTVictim=LLaDA 1.52026.0601
DIRECTVictim=LLaDA2.0-mini2026.0601
DIRECTVictim=Dream-Instruct2026.0601
PADVictim=LLaDA2.0-mini2026.0692.96
DIJAVictim=LLaDA2.0-mini2026.06153.25
CFATarget Model=Gemini 2.5 Pro2026.0426.74.22
CFATarget Model=GPT-4o2026.0430.63.82
PAIRTarget Model=GPT-4o2026.0439.84.39
PAIRTarget Model=Gemini 2.5 Pro2026.0440.84.79
CFATarget Model=GLM-4.52026.0444.64.75
PAIRTarget Model=Deepseek V32026.04455.01
PAIRTarget Model=GLM-4.52026.0445.65
MASKFORGE-GoalVictim=LLaDA2.0-mini, k=5, Pattern source=JailbreakBench/HarmBench/StrongReject, Retrieval strategy=Goal sampling2026.06474.57
PAIRTarget Model=Qwen 32026.0449.85.1
CIATarget Model=GPT-4o2026.0451.74.98
DIJAVictim=LLaDA 1.52026.06554.17
PADVictim=LLaDA-Instruct2026.06564.22
DIJAVictim=LLaDA-Instruct2026.06584.18
CrescendoTarget Model=GPT-4o2026.0458.35.76
CFATarget Model=Qwen 32026.0458.36
PADVictim=LLaDA 1.52026.06594.2
CFATarget Model=Deepseek V32026.04606.39
CIATarget Model=GLM-4.52026.0460.25.77
CIATarget Model=Qwen 32026.0460.25.31
CIATarget Model=Gemini 2.5 Pro2026.0460.86.22
MASKFORGE-GoalVictim=LLaDA 1.5, k=5, Pattern source=JailbreakBench/HarmBench/StrongReject, Retrieval strategy=Goal sampling2026.06614.69
DIJAVictim=Dream-Instruct2026.06624.28
DIRECTVictim=MMaDA-CoT2026.06624.23
DIJAVictim=MMaDA-CoT2026.06634.37
PADVictim=Dream-Instruct2026.06644.34
GOATTarget Model=GPT-4o2026.0465.36.01
CrescendoTarget Model=GLM-4.52026.0469.46.43
MASKFORGE-GoalVictim=Dream-Instruct, k=5, Pattern source=JailbreakBench/HarmBench/StrongReject, Retrieval strategy=Goal sampling2026.06704.37
GOATTarget Model=Gemini 2.5 Pro2026.0470.56.97
CIATarget Model=Deepseek V32026.0471.56.96
CrescendoTarget Model=Gemini 2.5 Pro2026.0473.57.05
MASKFORGE-GoalVictim=LLaDA-Instruct, k=5, Pattern source=JailbreakBench/HarmBench/StrongReject, Retrieval strategy=Goal sampling2026.06744.78
MASKFORGE-EmbedVictim=LLaDA2.0-mini, k=5, Pattern source=JailbreakBench/HarmBench/StrongReject, Retrieval strategy=Embedding similarity2026.06744.89
PADVictim=MMaDA-CoT2026.06764.66
CrescendoTarget Model=Deepseek V32026.0476.27.2
GOATTarget Model=GLM-4.52026.0479.47.49
CrescendoTarget Model=Qwen 32026.04807.57
GOATTarget Model=Deepseek V32026.04817.86
GOATTarget Model=Qwen 32026.0483.98.17
Salami AttackTarget Model=GPT-4o, Shots=5-shot2026.0486.98.47
MASKFORGE-EmbedVictim=LLaDA-Instruct, k=5, Pattern source=JailbreakBench/HarmBench/StrongReject, Retrieval strategy=Embedding similarity2026.06884.87
MASKFORGE-EmbedVictim=Dream-Instruct, k=5, Pattern source=JailbreakBench/HarmBench/StrongReject, Retrieval strategy=Embedding similarity2026.06904.81
Salami AttackTarget Model=GPT-4o, Shots=10-shot2026.0490.49.19
MASKFORGE-EmbedVictim=LLaDA 1.5, k=5, Pattern source=JailbreakBench/HarmBench/StrongReject, Retrieval strategy=Embedding similarity2026.06914.84
Salami AttackTarget Model=Gemini 2.5 Pro, Shots=5-shot2026.0491.29.06
Salami AttackTarget Model=GLM-4.5, Shots=5-shot2026.0491.38.98
Salami AttackTarget Model=Gemini 2.5 Pro, Shots=10-shot2026.0491.59.11
Salami AttackTarget Model=GLM-4.5, Shots=10-shot2026.0493.39.16
A-SalamiTarget Model=Gemini 2.5 Pro2026.0494.69.28
MASKFORGE-GoalVictim=MMaDA-CoT, k=5, Pattern source=JailbreakBench/HarmBench/StrongReject, Retrieval strategy=Goal sampling2026.06954.96
Salami AttackTarget Model=Qwen 3, Shots=5-shot2026.0495.29.36
A-SalamiTarget Model=GPT-4o2026.0495.89.37
Salami AttackTarget Model=Deepseek V3, Shots=5-shot2026.04969.24
Salami AttackTarget Model=Deepseek V3, Shots=10-shot2026.0496.79.44
Salami AttackTarget Model=Qwen 3, Shots=10-shot2026.0497.79.64
MASKFORGE-EmbedVictim=MMaDA-CoT, k=5, Pattern source=JailbreakBench/HarmBench/StrongReject, Retrieval strategy=Embedding similarity2026.06985
A-SalamiTarget Model=Deepseek V32026.0498.19.7
A-SalamiTarget Model=GLM-4.52026.0498.89.69
A-SalamiTarget Model=Qwen 32026.0499.29.71