Cascade Attack Detection on LLM Debate
84.2TPR@5%CASPIAN
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| CASPIAN2026.05 | 84.2 | 75.3 | 0.915 | |
| Full TraceCategory=LLM Judges, Evaluation Mode=offline2026.05 | 68.6 | — | 0.765 | |
| BlindGuardCategory=Graph-Based2026.05 | 61.2 | 46.8 | 0.802 | |
| GUARDIANCategory=Graph-Based2026.05 | 59.8 | 45.2 | 0.796 | |
| G-SafeguardCategory=Graph-Based2026.05 | 58.2 | 43.7 | 0.785 | |
| Sliding WindowCategory=LLM Judges, Window size (W)=102026.05 | 57.5 | 39.5 | 0.736 | |
| Sliding WindowCategory=LLM Judges, Window size (W)=52026.05 | 49.6 | 31.4 | 0.717 | |
| Single TurnCategory=LLM Judges2026.05 | 38.3 | 19 | 0.661 | |
| PromptArmorCategory=Guardrails2026.05 | 30.4 | 20.7 | 0.622 | |
| JailGuardCategory=Guardrails2026.05 | 27.1 | 16.6 | 0.586 | |
| Prompt Guard 2Category=Guardrails2026.05 | 23.3 | 15.4 | 0.579 | |
| PerplexityCategory=Distributional2026.05 | 8.6 | 4.5 | 0.533 |