https://xinjie-shen.com/tag/jailbreak/JailbreakWowchemy (https://wowchemy.com)en-usSat, 01 Nov 2025 00:00:00 +0000https://xinjie-shen.com/media/icon_hu646f7301b7fde7528ecdae8cec89fc29_9606_512x512_fill_lanczos_center_3.pnghttps://xinjie-shen.com/tag/jailbreak/https://xinjie-shen.com/publication/preprint/cka-agent/Sat, 01 Nov 2025 00:00:00 +0000https://xinjie-shen.com/publication/preprint/cka-agent/<h2 id="key-contributions">Key Contributions</h2> <ul> <li><strong>Novel Attack Paradigm</strong>: Reformulate jailbreaking from static prompt optimization to dynamic knowledge decomposition and adaptive tree search</li> <li><strong>High Success Rates</strong>: Achieve 96-99% attack success rates against GPT, Gemini, and Claude models</li> <li><strong>Efficiency</strong>: 70-95% first-iteration success with 92-95% success within two iterations</li> <li><strong>Defense Insights</strong>: Reveal that current guardrails cannot aggregate intent across adaptively constructed innocuous queries</li> </ul> <h2 id="methodology">Methodology</h2> <p>CKA-Agent operates through iterative exploration cycles:</p> <ol> <li>Generate locally harmless sub-queries that extract correlated knowledge</li> <li>Execute queries against the target and collect responses</li> <li>Evaluate responses using hybrid scoring (introspection + target feedback)</li> <li>Branch adaptively based on UCT-guided selection</li> <li>Synthesize accumulated knowledge from successful trajectories</li> <li>Backpropagate failure signals to guide future iterations</li> </ol>