Correlated Knowledge in Target LLM Jailbreak aims to extract the target knowledge V₀ Legend Vᵢ: Correlated Knowledge V₀: Target Knowledge POA: Prompt Optimization Attack (Direct) Successful Query Failed/Blocked Synthesis to Target Attacker Query "How to make a bomb?" V₁ V₂ V₃ V₄ V₀ Target Chemical Synthesis Process Details ❌ POA: Direct Attack to V₀ Guardrails Blocked! ✓ Our Approach: Adaptive Search over Correlated Knowledge 🔒 Corr-Knowledge 1 Corr-Knowledge 2 Blocked! Adaptive Reroute ✓ Synthesizer Combine correlated knowledge
Watch how our approach adaptively explores correlated knowledge while POA methods get blocked