List your company
ResearchFormerly WizWand

Multimodal Jailbreak Attack on HarmBench

0ASRFigStep
-3.1418.05539.2560.445Oct 20, 2025Nov 7, 2025Nov 25, 2025Dec 13, 2025Dec 31, 2025Jan 18, 2026Feb 5, 2026
Updated 5mo ago

Evaluation results

MethodLinks
FigStepEvaluation Model=HarmBench, Target Model=GPT-4o2025.100
FigStepEvaluation Model=GPT-4o-mini, Target Model=GPT-4o2025.100
FigStepEvaluation Model=Average, Target Model=GPT-4o2025.100
HADESEvaluation Model=HarmBench, Target Model=GPT-4o-mini2025.103.5
HADESEvaluation Model=HarmBench, Target Model=GPT-4o2025.103.5
HADESEvaluation Model=Average, Target Model=GPT-4o-mini2025.103.75
HADESEvaluation Model=GPT-4o-mini, Target Model=GPT-4o-mini2025.104
HADESEvaluation Model=Average, Target Model=GPT-4o2025.104
HADESEvaluation Model=GPT-4o-mini, Target Model=GPT-4o2025.104.5
FigStepEvaluation Model=GPT-4o-mini, Target Model=GPT-4o-mini2025.108
FigStepEvaluation Model=Average, Target Model=GPT-4o-mini2025.109
CS-DJEvaluation Model=HarmBench, Target Model=GPT-4o2025.109.5
Direct RequestType=Single-turn, Target Model=Sonnet 4.52026.0210
FigStepEvaluation Model=HarmBench, Target Model=GPT-4o-mini2025.1010
FigStepEvaluation Model=HarmBench, Target Model=Qwen2.5-VL-7B2025.1013
Direct PlanType=Multi-turn, Target Model=Sonnet 4.52026.0213.3
CS-DJEvaluation Model=Average, Target Model=GPT-4o2025.1014
Direct RequestType=Single-turn, Target Model=Qwen3-VL-8B2026.0216.7
FigStepType=Single-turn, Target Model=Sonnet 4.52026.0216.7
SI-AttackType=Single-turn, Target Model=Sonnet 4.52026.0216.7
CS-DJEvaluation Model=GPT-4o-mini, Target Model=GPT-4o2025.1018.5
SSAType=Multi-turn, Target Model=Sonnet 4.52026.0220
CS-DJEvaluation Model=HarmBench, Target Model=GPT-4o-mini2025.1020.5
FigStepEvaluation Model=Average, Target Model=Qwen2.5-VL-7B2025.1021.5
Direct PlanType=Multi-turn, Target Model=Qwen3-VL-8B2026.0223.3
CrescendoType=Multi-turn, Target Model=Sonnet 4.52026.0223.3
SI-AttackType=Single-turn, Target Model=Qwen3-VL-8B2026.0230
FigStepEvaluation Model=GPT-4o-mini, Target Model=Qwen2.5-VL-7B2025.1030
CS-DJEvaluation Model=Average, Target Model=GPT-4o-mini2025.1030.75
FigStepType=Single-turn, Target Model=Qwen3-VL-8B2026.0233.3
SSAType=Multi-turn, Target Model=Qwen3-VL-8B2026.0233.3
MM-PlanType=Multi-turn, Target Model=Sonnet 4.52026.0236.7
CrescendoType=Multi-turn, Target Model=Qwen3-VL-8B2026.0240
CS-DJEvaluation Model=GPT-4o-mini, Target Model=GPT-4o-mini2025.1041
HADESEvaluation Model=HarmBench, Target Model=Qwen2.5-VL-7B2025.1045.5
HADESEvaluation Model=Average, Target Model=Qwen2.5-VL-7B2025.1046.75
HADESEvaluation Model=GPT-4o-mini, Target Model=Qwen2.5-VL-7B2025.1048
HADESEvaluation Model=HarmBench, Target Model=InternVL3-8B2025.1050.5
CS-DJEvaluation Model=HarmBench, Target Model=Qwen2.5-VL-7B2025.1050.5
HADESEvaluation Model=Average, Target Model=InternVL3-8B2025.1051.5
HADESEvaluation Model=GPT-4o-mini, Target Model=InternVL3-8B2025.1052.5
CS-DJEvaluation Model=Average, Target Model=Qwen2.5-VL-7B2025.1053
CS-DJEvaluation Model=HarmBench, Target Model=InternVL3-8B2025.1054
CS-DJEvaluation Model=GPT-4o-mini, Target Model=Qwen2.5-VL-7B2025.1055.5
FigStepEvaluation Model=HarmBench, Target Model=InternVL3-8B2025.1058.5
CS-DJEvaluation Model=Average, Target Model=InternVL3-8B2025.1059.5
FigStepEvaluation Model=Average, Target Model=InternVL3-8B2025.1059.75
VERA-VEvaluation Model=HarmBench, Target Model=GPT-4o-mini2025.1060
VERA-VEvaluation Model=Average, Target Model=GPT-4o-mini2025.1060.5
FigStepEvaluation Model=GPT-4o-mini, Target Model=InternVL3-8B2025.1061
VERA-VEvaluation Model=GPT-4o-mini, Target Model=GPT-4o-mini2025.1061
MM-PlanType=Multi-turn, Target Model=Qwen3-VL-8B2026.0263.3
CS-DJEvaluation Model=GPT-4o-mini, Target Model=InternVL3-8B2025.1065
VERA-VEvaluation Model=HarmBench, Target Model=GPT-4o2025.1065
VERA-VEvaluation Model=Average, Target Model=GPT-4o2025.1067.75
VERA-VEvaluation Model=GPT-4o-mini, Target Model=GPT-4o2025.1070.5
VERA-VEvaluation Model=GPT-4o-mini, Target Model=Qwen2.5-VL-7B2025.1071
VERA-VEvaluation Model=Average, Target Model=Qwen2.5-VL-7B2025.1072
VERA-VEvaluation Model=HarmBench, Target Model=Qwen2.5-VL-7B2025.1073
VERA-VEvaluation Model=HarmBench, Target Model=InternVL3-8B2025.1074.5
VERA-VEvaluation Model=Average, Target Model=InternVL3-8B2025.1076.5
VERA-VEvaluation Model=GPT-4o-mini, Target Model=InternVL3-8B2025.1078.5