Jailbreak Attack on Harmful Query Evaluation Set GPT-5.4-mini N=750
1.01Toxicity ScoreFigStep
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| FigStepNumber of Trials (T)=12026.05 | 1.01 | 0.61 | 0.27 | 0.53 | |
| HADESNumber of Trials (T)=12026.05 | 1.02 | 0.49 | 0.67 | 1.73 | |
| FlipAttackNumber of Trials (T)=12026.05 | 1.04 | 0.66 | 0.53 | 1.87 | |
| SINumber of Trials (T)=52026.05 | 1.21 | 1.32 | 5.6 | 6 | |
| CS-DJNumber of Trials (T)=12026.05 | 1.4 | 1.83 | 11.2 | 7.6 | |
| TxtV-GDINumber of Trials (T)=52026.05 | 3.24 | 4.72 | 58.4 | 53.33 | |
| TypVNumber of Trials (T)=52026.05 | 3.39 | 5.31 | 61.87 | 61.07 | |
| TxtVNumber of Trials (T)=52026.05 | 3.46 | 5.23 | 63.47 | 63.87 | |
| TTVNumber of Trials (T)=52026.05 | 3.8 | 5.84 | 73.2 | 69.47 | |
| TTV-GDINumber of Trials (T)=5, Kd=62026.05 | 4.08 | 6.3 | 79.6 | 78.8 |