Jailbreak Attack on Harmful Query Evaluation Set Claude Haiku 4.5 N=750
1.01ToxicityHADES
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| HADESNumber of Trials (T)=12026.05 | 1.01 | 0.06 | 0.67 | 1.47 | |
| FlipAttackNumber of Trials (T)=12026.05 | 1.01 | 0.11 | 0.53 | 1.07 | |
| FigStepNumber of Trials (T)=12026.05 | 1.03 | 0.08 | 0.53 | 1.33 | |
| SINumber of Trials (T)=52026.05 | 1.15 | 0.51 | 3.73 | 4.67 | |
| CS-DJNumber of Trials (T)=12026.05 | 1.28 | 0.8 | 6.8 | 5.73 | |
| TxtV-GDINumber of Trials (T)=52026.05 | 1.54 | 1.34 | 14.67 | 14.53 | |
| TxtVNumber of Trials (T)=52026.05 | 2.03 | 2.48 | 27.2 | 25.33 | |
| TTV-GDINumber of Trials (T)=5, Kd=62026.05 | 2.39 | 3.21 | 35.6 | 34.67 | |
| TypVNumber of Trials (T)=52026.05 | 2.42 | 3.23 | 37.87 | 38 | |
| TTVNumber of Trials (T)=52026.05 | 2.54 | 3.46 | 39.73 | 39.33 |