Jailbreaking on AdvBench (WASR & W-Toxicity Score)
73.3WASRwide-net-casting jailbreak method (Ours)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| wide-net-casting jailbreak method (Ours)Safety Alignment=Original Safety Alignment2026.05 | 73.3 | 67.2 | |
| Naive Strategy 2Safety Alignment=Original Safety Alignment2026.05 | 56.2 | 52 | |
| Naive Strategy 1Safety Alignment=Original Safety Alignment2026.05 | 55.5 | 51.3 | |
| wide-net-casting jailbreak method (Ours)Safety Alignment=Original Safety Alignment + SmoothLLM2026.05 | 51.9 | 46.2 | |
| ReMissSafety Alignment=Original Safety Alignment2026.05 | 50.2 | 46.9 | |
| wide-net-casting jailbreak method (Ours)Safety Alignment=Original Safety Alignment + RobustKV2026.05 | 40.6 | 36.3 | |
| Naive Strategy 2Safety Alignment=Original Safety Alignment + SmoothLLM2026.05 | 38 | 33.6 | |
| Naive Strategy 1Safety Alignment=Original Safety Alignment + SmoothLLM2026.05 | 37.3 | 32.5 | |
| ReMissSafety Alignment=Original Safety Alignment + SmoothLLM2026.05 | 33.1 | 29.3 | |
| Naive Strategy 2Safety Alignment=Original Safety Alignment + RobustKV2026.05 | 31.8 | 28.4 | |
| Naive Strategy 1Safety Alignment=Original Safety Alignment + RobustKV2026.05 | 31.1 | 27.9 | |
| ReMissSafety Alignment=Original Safety Alignment + RobustKV2026.05 | 28.5 | 25.7 |