Revolutionizing Large Language Model Safety with Causal Analysis

This research introduces a novel framework, Causal Analyst, to understand and mitigate "jailbreak" attacks on Large Language Models (LLMs). By integrating Generative AI with data-driven causal discovery, the work aims to fortify the safety and reliability of LLMs, paving the way for more secure and trustworthy AI systems.

• •Causal Analyst uses Generative AI to pinpoint causes of LLM jailbreaks.
• •The research identified specific prompt features (like "Positive Character") that directly cause jailbreaks.
• •The findings are applied to improve attack success and create more robust guardrails.