List your company

Jailbreak Attack on HADES (Harm Category Success Rates)

66.67Success Rate (Animal)CS-DJ
0.1117.3934.6751.95Feb 3, 2025Apr 4, 2025Jun 3, 2025Aug 3, 2025Oct 2, 2025Dec 1, 2025Jan 31, 2026

Evaluation results

MethodLinks
CS-DJVictim Model=Qwen3-VL-30B-A3B2026.0166.6768.6749.3332.676456.27
Text-DJVictim Model=Qwen3-VL-30B-A3B2026.0163.3370.6748.6738.6766.6757.6
Text-DJVictim Model=Qwen3-VL-4B2026.015661.334231.335048.13
CS-DJVictim Model=Qwen3-VL-8B2026.0149.3349.3332.672041.3338.53
CS-DJVictim Model=Qwen3-VL-4B2026.0147.3341.33301835.3334.4
CLIP2025.024459.6769.22—77.2262.41
Text-DJVictim Model=Qwen3-VL-8B2026.0142.6770.6760266452.67
FARE^4Robustness Budget=4/2552025.0233.6752.4465.33—72.3355.94
CS-DJVictim Model=GPT 4.1 mini2026.0133.3374.6777.333064.6756
Text-DJVictim Model=GPT 4o mini2026.013279.3378.673868.6759.33
Sim-CLIP^4Robustness Budget=4/2552025.0231.8952.4466.11—72.6755.78
Robust-LLaVA^4_GRobustness Budget=4/2552025.0230.224353.89—65.1148.06
Text-DJVictim Model=GPT 4.1 mini2026.0128.6785.3380.6745.3381.3364.27
Robust-LLaVA^4_HRobustness Budget=4/2552025.0227.5638.1151.11—64.8945.41
CS-DJVictim Model=GPT 4o mini2026.01246466.6720.6758.6746.8
CS-DJVictim Model=Gemini-2.5-Flash2026.0116.6764.6744.6726.675040.53
Text-DJVictim Model=Gemini-2.5-Flash2026.011673.335428.674844
HADESVictim Model=Gemini-2.5-Flash2026.011039.3329.334.6723.3321.33
HADESVictim Model=Qwen3-VL-4B2026.017.3318.677.3322411.87
HADESVictim Model=Qwen3-VL-30B-A3B2026.017.332615.33613.3313.6
HADESVictim Model=GPT 4.1 mini2026.01414123.338.678.4
HADESVictim Model=Qwen3-VL-8B2026.013.3314.677.333.33189.33
HADESVictim Model=GPT 4o mini2026.012.67681.332.674.13