One Turn Too Late: Response-Aware Defense Against Hidden Malicious Intent in Multi-Turn Dialogue
A response-aware monitor trained to intervene at the first harm-enabling turn while preserving benign exploratory dialogue.
Xinjie Shen
LLM Safety
The Trojan Knowledge: Bypassing Commercial LLM Guardrails via Harmless Prompt Weaving and Adaptive Tree Search
Introduce CKA-Agent, a framework that reformulates jailbreaking as an adaptive tree search over the target LLM’s correlated knowledge, achieving 96-99% attack success rates against state-of-the-art commercial LLMs.
New Preprint: CKA-Agent Achieves 96-99% Jailbreak Success on GPT, Gemini & Claude
Our new paper introduces CKA-Agent, a novel framework that bypasses commercial LLM guardrails through adaptive tree search over correlated knowledge.