This research reveals a novel jailbreak attack method that exploits LLMs' analytical reasoning to bypass safety guardrails with high effectiveness.

• Introduces ABJ (Analyzing-based Jailbreak) attack that leverages reasoning capabilities inherent to LLMs
• Achieves significantly higher success rates than previous jailbreak methods
• Demonstrates attack effectiveness across multiple commercial models including ChatGPT and Claude
• Shows that current safety measures often fail against attacks mimicking analytical reasoning processes

This work exposes critical security implications for LLM deployment in enterprise environments, highlighting the urgent need for more robust defense mechanisms against reasoning-based attacks.

LLMs can be Dangerous Reasoners: Analyzing-based Jailbreak Attack on Large Language Models