List your company

Cascade Attack Detection on LLM Debate

84.2TPR@5%CASPIAN
5.57625.98846.466.812May 19, 2026

Evaluation results

MethodLinks
CASPIAN2026.0584.275.30.915
Full TraceCategory=LLM Judges, Evaluation Mode=offline2026.0568.6—0.765
BlindGuardCategory=Graph-Based2026.0561.246.80.802
GUARDIANCategory=Graph-Based2026.0559.845.20.796
G-SafeguardCategory=Graph-Based2026.0558.243.70.785
Sliding WindowCategory=LLM Judges, Window size (W)=102026.0557.539.50.736
Sliding WindowCategory=LLM Judges, Window size (W)=52026.0549.631.40.717
Single TurnCategory=LLM Judges2026.0538.3190.661
PromptArmorCategory=Guardrails2026.0530.420.70.622
JailGuardCategory=Guardrails2026.0527.116.60.586
Prompt Guard 2Category=Guardrails2026.0523.315.40.579
PerplexityCategory=Distributional2026.058.64.50.533