Jailbreaking MLLMs on AdvBench
100WASRwide-net-casting jailbreak method
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| wide-net-casting jailbreak methodSafety Alignment=Original Safety Alignment2026.05 | 100 | 0.94 | |
| Ours (Wide-Net-Casting)Safety Alignment Strategy=Original Safety Alignment2026.05 | 99.6 | 0.932 | |
| Naive Strategy 2Safety Alignment=Original Safety Alignment2026.05 | 95.8 | 0.898 | |
| Naive Strategy 1Safety Alignment=Original Safety Alignment2026.05 | 95.5 | 0.883 | |
| Naive Strategy 2Safety Alignment Strategy=Original Safety Alignment2026.05 | 93.6 | 0.897 | |
| MLAI+PixArt-αSafety Alignment=Original Safety Alignment2026.05 | 93.3 | 0.867 | |
| Naive Strategy 1Safety Alignment Strategy=Original Safety Alignment2026.05 | 92.4 | 0.886 | |
| Baseline (MLAI + PixArt-α)Safety Alignment Strategy=Original Safety Alignment2026.05 | 89.4 | 0.858 | |
| wide-net-casting jailbreak methodSafety Alignment=Original Safety Alignment + VLGuard2026.05 | 50.8 | 0.473 | |
| wide-net-casting jailbreak methodSafety Alignment=Original Safety Alignment + IMMUNE2026.05 | 47.8 | 0.44 | |
| Ours (Wide-Net-Casting)Safety Alignment Strategy=Original Safety Alignment + VLGuard2026.05 | 46.2 | 0.412 | |
| Ours (Wide-Net-Casting)Safety Alignment Strategy=Original Safety Alignment + IMMUNE2026.05 | 44.2 | 0.406 | |
| wide-net-casting jailbreak methodSafety Alignment=Original Safety Alignment + ASTRA2026.05 | 42 | 0.387 | |
| Naive Strategy 2Safety Alignment=Original Safety Alignment + VLGuard2026.05 | 41.1 | 0.363 | |
| Ours (Wide-Net-Casting)Safety Alignment Strategy=Original Safety Alignment + ASTRA2026.05 | 40.6 | 0.365 | |
| Naive Strategy 1Safety Alignment=Original Safety Alignment + VLGuard2026.05 | 40.6 | 0.355 | |
| Naive Strategy 2Safety Alignment=Original Safety Alignment + IMMUNE2026.05 | 39.4 | 0.351 | |
| Naive Strategy 1Safety Alignment=Original Safety Alignment + IMMUNE2026.05 | 38.6 | 0.344 | |
| MLAI+PixArt-αSafety Alignment=Original Safety Alignment + VLGuard2026.05 | 37.5 | 0.311 | |
| MLAI+PixArt-αSafety Alignment=Original Safety Alignment + IMMUNE2026.05 | 36.9 | 0.32 | |
| Naive Strategy 2Safety Alignment Strategy=Original Safety Alignment + VLGuard2026.05 | 35.5 | 0.318 | |
| Naive Strategy 1Safety Alignment Strategy=Original Safety Alignment + VLGuard2026.05 | 34.7 | 0.317 | |
| Naive Strategy 2Safety Alignment Strategy=Original Safety Alignment + IMMUNE2026.05 | 34.4 | 0.307 | |
| Naive Strategy 2Safety Alignment=Original Safety Alignment + ASTRA2026.05 | 33.9 | 0.291 | |
| Naive Strategy 1Safety Alignment Strategy=Original Safety Alignment + IMMUNE2026.05 | 33.8 | 0.295 | |
| Naive Strategy 1Safety Alignment=Original Safety Alignment + ASTRA2026.05 | 33.2 | 0.277 | |
| Naive Strategy 2Safety Alignment Strategy=Original Safety Alignment + ASTRA2026.05 | 32.7 | 0.292 | |
| Baseline (MLAI + PixArt-α)Safety Alignment Strategy=Original Safety Alignment + VLGuard2026.05 | 32.3 | 0.284 | |
| Naive Strategy 1Safety Alignment Strategy=Original Safety Alignment + ASTRA2026.05 | 32 | 0.283 | |
| Baseline (MLAI + PixArt-α)Safety Alignment Strategy=Original Safety Alignment + IMMUNE2026.05 | 31.5 | 0.273 | |
| MLAI+PixArt-αSafety Alignment=Original Safety Alignment + ASTRA2026.05 | 30.7 | 0.253 | |
| Baseline (MLAI + PixArt-α)Safety Alignment Strategy=Original Safety Alignment + ASTRA2026.05 | 28.1 | 0.242 |