List your company

Jailbreaking Attack on MM-SafetyBench

95Attack Success Rate (ASR)TAO-Attack
-3.404822.142647.6973.2374Jan 8, 2026Feb 6, 2026Mar 7, 2026Apr 6, 2026May 5, 2026Jun 3, 2026Jul 3, 2026

Evaluation results

MethodLinks
TAO-AttackTarget Model=Qwen2.5-VL-7B-Instruct2026.0395132
DMNTarget Model=Gemini-2.5-pro2026.0593.75—
DMNTarget Model=Claude Sonnet 42026.0593.1—
Multi-turn AttackModel=LLaVa-7B, Turn=32026.0191.5—
DMNTarget Model=GPT-4V2026.0591.25—
INFER+Target Model=InternVL2, Model Type=Open-source2026.0790.68—
DMNTarget Model=GPT-4o2026.0590.42—
I-GCGTarget Model=Qwen2.5-VL-7B-Instruct2026.0390248
DMNTarget Model=Seed1.5-VL2026.0588.45—
INFER+Target Model=Llama 3.2-Vision, Model Type=Open-source2026.0787.41—
INFER+Target Model=Qwen2-VL, Model Type=Open-source2026.0786.25—
DMNTarget Model=Average2026.0586.24—
INFER+Target Model=InternVL3.5, Model Type=Open-source2026.0783.92—
DMNTarget Model=GPT-52026.0582.86—
INFER+Target Model=GPT-4.1 mini, Model Type=Closed-source, Surrogate Model=Qwen2-VL2026.0782.75—
DMNTarget Model=o32026.0582.02—
INFERTarget Model=InternVL2, Model Type=Open-source2026.0781—
DMNTarget Model=o4-mini2026.0580.65—
DMNTarget Model=Gemini-2.5-flash2026.0580.54—
DMNTarget Model=Qwen3-VL-plus2026.0579.4—
INFERTarget Model=Qwen2-VL, Model Type=Open-source2026.0779.02—
INFER+Target Model=Gemini 2.5 Flash Lite, Model Type=Closed-source, Surrogate Model=Qwen2-VL2026.0778.79—
Multi-turn AttackModel=GPT-4o, Turn=32026.0177.31—
INFER+Target Model=Gemini 2.5 Flash, Model Type=Closed-source, Surrogate Model=Qwen2-VL2026.0776.57—
INFERTarget Model=Llama 3.2-Vision, Model Type=Open-source2026.0776.46—
CS-DJTarget Model=Qwen2-VL, Model Type=Open-source2026.0775.06—
INFERTarget Model=Gemini 2.5 Flash Lite, Model Type=Closed-source2026.0774.83—
INFER+Target Model=Qwen3-VL, Model Type=Open-source2026.0774.48—
CS-DJTarget Model=InternVL2, Model Type=Open-source2026.0772.26—
MM-SafetyBenchModel=LLaVa-7B2026.0172.14—
CS-DJTarget Model=InternVL3.5, Model Type=Open-source2026.0769.7—
SI-AttackModel=GPT-4o2026.0168.57—
INFERTarget Model=InternVL3.5, Model Type=Open-source2026.0763.4—
SI-AttackModel=LLaVa-7B2026.0162.68—
FigStepModel=LLaVa-7B2026.0162.3—
INFERTarget Model=Gemini 2.5 Flash, Model Type=Closed-source2026.0757.81—
CS-DJTarget Model=GPT-4.1 mini, Model Type=Closed-source2026.0756.41—
CS-DJTarget Model=Gemini 2.5 Flash Lite, Model Type=Closed-source2026.0756.18—
CS-DJTarget Model=Gemini 2.5 Flash, Model Type=Closed-source2026.0754.66—
INFERTarget Model=GPT-4.1 mini, Model Type=Closed-source2026.0754.19—
ImageTypoTarget Model=InternVL2, Model Type=Open-source2026.0753.38—
TVPITarget Model=Qwen2-VL, Model Type=Open-source2026.0753.26—
TVPITarget Model=InternVL2, Model Type=Open-source2026.0753.26—
ImageTypoTarget Model=Qwen2-VL, Model Type=Open-source2026.0751.75—
TVPITarget Model=Qwen3-VL, Model Type=Open-source2026.0745.22—
INFERTarget Model=Qwen3-VL, Model Type=Open-source2026.0744.17—
CS-DJTarget Model=Llama 3.2-Vision, Model Type=Open-source2026.0739.86—
TVPITarget Model=Llama 3.2-Vision, Model Type=Open-source2026.0738.23—
CS-DJTarget Model=Qwen3-VL, Model Type=Open-source2026.0733.45—
ImageTypoTarget Model=Gemini 2.5 Flash, Model Type=Closed-source2026.0731—
ImageTypoTarget Model=Llama 3.2-Vision, Model Type=Open-source2026.0730.42—
QRATarget Model=GPT-4o2026.0529.4—
QRATarget Model=Gemini-2.5-flash2026.0528.93—
QRATarget Model=GPT-4V2026.0528.27—
TVPITarget Model=Gemini 2.5 Flash, Model Type=Closed-source2026.0727.51—
QRATarget Model=Seed1.5-VL2026.0526.9—
QRATarget Model=Qwen3-VL-plus2026.0525.48—
QRATarget Model=Average2026.0521.3—
TVPITarget Model=Gemini 2.5 Flash Lite, Model Type=Closed-source2026.0720.98—
FigStepModel=GPT-4o2026.0119.81—
ImageTypoTarget Model=InternVL3.5, Model Type=Open-source2026.0719.7—
QRATarget Model=Gemini-2.5-pro2026.0519.29—
TVPITarget Model=InternVL3.5, Model Type=Open-source2026.0718.53—
QRATarget Model=o4-mini2026.0516.49—
QRATarget Model=Claude Sonnet 42026.0515.24—
ImageTypoTarget Model=Gemini 2.5 Flash Lite, Model Type=Closed-source2026.0714.8—
ImageTypoTarget Model=GPT-4.1 mini, Model Type=Closed-source2026.0712.59—
MM-SafetyBenchModel=GPT-4o2026.0112.34—
QRATarget Model=GPT-52026.0511.73—
QRATarget Model=o32026.0511.31—
TVPITarget Model=GPT-4.1 mini, Model Type=Closed-source2026.0710.37—
Mistral-Medium-32026.0410—
LLaVA-v1.6-Vicuna-7BBackbone=Vicuna-7B2026.047.31—
ImageTypoTarget Model=Qwen3-VL, Model Type=Open-source2026.074.43—
Kimi-K2.52026.042.69—
Llama-3.2-11B-VisionParameters=11B, Type=Vision2026.042.69—
Gemini-3-Flash2026.041.54—
Qwen3-VL-Plus2026.040.77—
Claude-Haiku-4.52026.040.77—
GPT-5-Mini2026.040.77—
DeepSeek-V3.22026.040.77—
Qwen3-VL-8B-InstructParameters=8B, Alignment=Instruct2026.040.38—