Adversarial Attack on LibrAI Do-Not-Answer and IndicAlign (145 prompts)
59ASR (%)GCG (Brute-Force)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GCG (Brute-Force)Target Model=Vicuna 7B2026.01 | 59 | 6.3 | |
| PEZTarget Model=Vicuna 7B2026.01 | 39 | 7.3 | |
| GBDATarget Model=Vicuna 7B2026.01 | 35 | 7.1 | |
| RECAPTarget Model=Vicuna 7B2026.01 | 33 | 4 |