This research introduces GraphAttack, a novel methodology for systematically generating jailbreak prompts that bypass LLM safety mechanisms through semantic transformations.

• Uses graph structures to represent malicious prompts with edges denoting different transformations
• Leverages Abstract Meaning Representation (AMR) and Resource Description Framework (RDF) to create semantically equivalent but undetectable harmful prompts
• Demonstrates systematic exploitation of representational blindspots in current LLM safety mechanisms
• Highlights critical security vulnerabilities that require immediate attention from AI developers

This work matters for security professionals because it exposes fundamental weaknesses in current AI safeguarding approaches, providing insights for developing more robust defense mechanisms against increasingly sophisticated prompt attacks.

GraphAttack: Exploiting Representational Blindspots in LLM Safety Mechanisms