This research introduces a novel approach to creating adversarial attacks against Large Language Models by leveraging mechanistic interpretability instead of relying solely on gradient computation.

Key Findings:

• Combines interpretability techniques with practical attack development
• Uses subspace rerouting to identify and exploit security vulnerabilities
• Bridges the gap between theoretical understanding and practical attack implementation
• Enables more precise targeting of LLM weaknesses

Security Implications: This work demonstrates how deeper understanding of LLM internal mechanisms creates more effective attack vectors, highlighting the need for defensive strategies that account for model internals rather than just surface-level behaviors.

Using Mechanistic Interpretability to Craft Adversarial Attacks against Large Language Models