Jailbreak Attack on HarmBench (first 200 harmful queries)
92.5ASR (%) (LLaMA-3-8B)Furina
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| FurinaCategory=Ours2026.05 | 92.5 | 94 | 90.5 | 93.5 | 83.5 | |
| ActorBreaker (ACL’25)Category=Multi-turn attacks2026.05 | 79 | 82 | 86 | 85.5 | 65 | |
| Crescendo (USENIX’25)Category=Multi-turn attacks2026.05 | 60 | 60 | 62 | 63.5 | 29.5 | |
| CodeAttack (ACL’24)Category=Single-turn attacks2026.05 | 46 | 69 | 71 | 59.5 | 21.5 | |
| AutoDAN-Turbo (ICLR’25)Category=Single-turn attacks2026.05 | 42.5 | 83 | 87 | 66 | 14 | |
| AmpleGCG (COLM’24)Category=Single-turn attacks2026.05 | 34.5 | 3 | 0 | 0 | 0 | |
| CoA (ACL’25)Category=Multi-turn attacks2026.05 | 25.5 | 25.5 | 18 | 27 | 13 | |
| PAIR (NeurIPS’23)Category=Single-turn attacks2026.05 | 18 | 1 | 2 | 0.5 | 0 | |
| CipherChat (ICLR’24)Category=Single-turn attacks2026.05 | 0 | 36 | 10 | 12.5 | 4 |