Jailbreak Attack on Harmful Query Evaluation Set Gemini-2.5-Flash N=750
2.18ToxicityHADES
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| HADESNumber of Trials (T)=12026.05 | 2.18 | 2.68 | 31.33 | 27.33 | |
| SINumber of Trials (T)=52026.05 | 3.02 | 4.44 | 50.8 | 54.93 | |
| CS-DJNumber of Trials (T)=12026.05 | 3.71 | 6.14 | 75.2 | 44.93 | |
| FigStepNumber of Trials (T)=12026.05 | 4.49 | 7.71 | 89.2 | 88.27 | |
| FlipAttackNumber of Trials (T)=12026.05 | 4.78 | 8.79 | 95.07 | 83.07 | |
| TypVNumber of Trials (T)=52026.05 | 4.83 | 8.88 | 96.93 | 97.73 | |
| TTVNumber of Trials (T)=52026.05 | 4.84 | 8.83 | 97.6 | 97.07 | |
| TxtV-GDINumber of Trials (T)=52026.05 | 4.85 | 9.11 | 97.47 | 97.6 | |
| TxtVNumber of Trials (T)=52026.05 | 4.88 | 9.08 | 97.73 | 98.13 | |
| TTV-GDINumber of Trials (T)=5, Kd=62026.05 | 4.88 | 8.97 | 97.73 | 97.2 |