Jailbreak attack on Safebench (test)
92IA ASRVRSA
Evaluation Results
| Method | Links | ||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| VRSAVictim Model=Qwen3-VL2025.12 | 92 | 4.56 | 56 | 3.6 | 76 | 4.2 | 74 | 3.8 | 60 | 3.74 | 58 | 3.84 | 76 | 4.12 | 70.29 | 3.98 | |
| MM-safetybenchVictim Model=LLaVA-NEXT2025.12 | 76 | 4.26 | 52 | 3.9 | 60 | 3.94 | 70 | 4.1 | 58 | 4.1 | 34 | 3.48 | 66 | 4.1 | 59.43 | 4.03 | |
| HADESVictim Model=Deepseek-VL22025.12 | 72 | 4.2 | 52 | 3.22 | 66 | 3.98 | 54 | 3.6 | 48 | 3.46 | 26 | 2.9 | 58 | 3.52 | 53.71 | 3.55 | |
| PRISMVictim Model=Deepseek-VL22025.12 | 70 | 4.2 | 72 | 4.3 | 66 | 3.9 | 84 | 4.5 | 58 | 3.7 | 66 | 3.9 | 64 | 3.9 | 68.57 | 4.08 | |
| VRSAVictim Model=Deepseek-VL22025.12 | 70 | 4.18 | 68 | 3.88 | 70 | 4.2 | 78 | 4.4 | 62 | 3.84 | 68 | 4.08 | 74 | 4.36 | 70 | 4.14 | |
| VRSAVictim Model=InternVL-3.52025.12 | 68 | 4.18 | 70 | 3.96 | 66 | 3.9 | 66 | 3.82 | 60 | 3.7 | 72 | 4.04 | 62 | 3.88 | 66.29 | 3.92 | |
| FigStepVictim Model=LLaVA-NEXT2025.12 | 68 | 3.84 | 64 | 3.9 | 70 | 4.1 | 72 | 4.1 | 60 | 3.92 | 60 | 3.8 | 84 | 4.58 | 68.29 | 4.04 | |
| HADESVictim Model=LLaVA-NEXT2025.12 | 66 | 3.76 | 66 | 3.84 | 84 | 4.36 | 68 | 3.76 | 68 | 3.88 | 44 | 3.54 | 74 | 3.74 | 67.14 | 3.84 | |
| PRISMVictim Model=LLaVA-NEXT2025.12 | 66 | 4 | 72 | 4.2 | 80 | 4.3 | 78 | 4.1 | 68 | 4 | 56 | 3.8 | 76 | 4.1 | 70.86 | 4.07 | |
| VRSAVictim Model=LLaVA-NEXT2025.12 | 64 | 3.88 | 74 | 4.16 | 84 | 4.4 | 88 | 4.44 | 68 | 4.26 | 70 | 3.96 | 88 | 4.54 | 76.57 | 4.24 | |
| PRISMVictim Model=Qwen3-VL2025.12 | 64 | 3.8 | 52 | 3.6 | 70 | 4.1 | 80 | 4.4 | 62 | 3.8 | 46 | 3.2 | 54 | 3.4 | 61.14 | 3.76 | |
| HADESVictim Model=InternVL-3.52025.12 | 52 | 3.3 | 34 | 2.64 | 36 | 2.66 | 26 | 2.26 | 32 | 2.82 | 30 | 2.94 | 30 | 2.3 | 34.29 | 2.7 | |
| PRISMVictim Model=InternVL-3.52025.12 | 50 | 3.6 | 48 | 3.4 | 50 | 3.8 | 54 | 3.6 | 52 | 3.5 | 48 | 3.2 | 48 | 3.7 | 50 | 3.6 | |
| MM-safetybenchVictim Model=Deepseek-VL22025.12 | 48 | 3.38 | 24 | 2.7 | 56 | 3.6 | 54 | 3.6 | 54 | 3.6 | 42 | 3.5 | 58 | 3.76 | 48 | 3.46 | |
| FigStepVictim Model=InternVL-3.52025.12 | 38 | 2.68 | 26 | 2.32 | 50 | 3.2 | 48 | 3.18 | 46 | 2.98 | 36 | 2.66 | 30 | 2.66 | 39.14 | 2.81 | |
| FigStepVictim Model=Deepseek-VL22025.12 | 22 | 3.16 | 72 | 4.14 | 56 | 3.8 | 70 | 4.02 | 64 | 4.04 | 36 | 3.36 | 70 | 4.06 | 55.71 | 3.79 | |
| HADESVictim Model=Qwen3-VL2025.12 | 14 | 1.64 | 22 | 2.06 | 10 | 1.42 | 34 | 2.44 | 18 | 1.96 | 24 | 2.36 | 22 | 2.14 | 20.57 | 2 | |
| FigStepVictim Model=Qwen3-VL2025.12 | 12 | 1.56 | 8 | 1.38 | 12 | 1.5 | 4 | 1.14 | 10 | 1.48 | 20 | 1.92 | 14 | 1.56 | 11.43 | 1.51 | |
| MM-safetybenchVictim Model=InternVL-3.52025.12 | 4 | 1.2 | 18 | 1.9 | 48 | 3.2 | 50 | 3.1 | 32 | 2.3 | 46 | 3.48 | 32 | 2.42 | 32.86 | 2.52 | |
| MM-safetybenchVictim Model=Qwen3-VL2025.12 | 0 | 1 | 10 | 1.8 | 16 | 1.8 | 32 | 2.4 | 2 | 1.4 | 6 | 1.6 | 4 | 1.28 | 10 | 1.61 |