Adversarial Attack on AdvBench (Specific LLM ASR)
18ASR (GPT-4o mini)Misrouter+FFA
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Misrouter+FFASetting=Black-box transfer setting, Surrogate Source=Surrogate MoE models2026.05 | 18 | 6 | 10 | 96 | 26 | 96 | |
| JailbrokenSetting=Black-box transfer setting, Surrogate Source=Surrogate MoE models2026.05 | 12 | 0 | 0 | 86 | 20 | 30 | |
| FFASetting=Black-box transfer setting, Surrogate Source=Surrogate MoE models2026.05 | 2 | 0 | 0 | 92 | 20 | 96 | |
| Direct InstructionSetting=Black-box transfer setting, Surrogate Source=Surrogate MoE models2026.05 | 0 | 0 | 0 | 6 | 2 | 2 | |
| GCGSetting=Black-box transfer setting, Surrogate Source=Surrogate MoE models2026.05 | 0 | 0 | 0 | 0 | 0 | 0 | |
| WildPromptSetting=Black-box transfer setting, Surrogate Source=Surrogate MoE models2026.05 | 0 | 4 | 0 | 96 | 22 | 32 | |
| MisrouterSetting=Black-box transfer setting, Surrogate Source=Surrogate MoE models2026.05 | 0 | 0 | 6 | 10 | 8 | 2 |