Jailbreak Attack Success Rate by Model on JailbreakBench
77Llama2 7B Attack Success RateLASH
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| LASHDefense=Perplexity2026.05 | 77 | 84 | 81 | 77 | 78 | 28 | 70.83 | |
| LASHDefense=Smooth LLM2026.05 | 47 | 58 | 43 | 33 | 57 | 11 | 41.5 | |
| TAPDefense=Perplexity2026.05 | 31 | 68 | 17 | 51 | 53 | 24 | 40.67 | |
| TAPDefense=Smooth LLM2026.05 | 22 | 38 | 11 | 24 | 44 | 10 | 24.83 | |
| LASHDefense=Llama Guard2026.05 | 13 | 17 | 17 | 18 | 26 | 17 | 18 | |
| FlipAttackDefense=Perplexity2026.05 | 8 | 51 | 30 | 77 | 17 | 10 | 32.17 | |
| TAPDefense=Llama Guard2026.05 | 7 | 9 | 5 | 7 | 10 | 11 | 8.17 | |
| AutoDAN-TDefense=Perplexity2026.05 | 7 | 44 | 30 | 37 | 57 | 11 | 31 | |
| AutoDANDefense=Llama Guard2026.05 | 4 | 1 | 1 | — | — | — | 2 | |
| AutoDANDefense=Smooth LLM2026.05 | 3 | 3 | 40 | — | — | — | 15.33 | |
| AutoDAN-TDefense=Smooth LLM2026.05 | 3 | 44 | 13 | 32 | 49 | 5 | 24.33 | |
| AutoDAN-TDefense=Llama Guard2026.05 | 2 | 39 | 2 | 18 | 41 | 9 | 18.5 | |
| PAIRDefense=Perplexity2026.05 | 1 | 18 | 49 | 40 | 50 | 1 | 26.5 | |
| PAIRDefense=Smooth LLM2026.05 | 0 | 4 | 5 | 3 | 28 | 0 | 6.67 | |
| FlipAttackDefense=Smooth LLM2026.05 | 0 | 4 | 0 | 36 | 0 | 0 | 6.67 | |
| PAIRDefense=Llama Guard2026.05 | 0 | 10 | 9 | 13 | 21 | 1 | 9 | |
| FlipAttackDefense=Llama Guard2026.05 | 0 | 4 | 2 | 4 | 4 | 5 | 3.17 | |
| AutoDANDefense=Perplexity2026.05 | 0 | 7 | 53 | — | — | — | 20 |