LLMmap introduces a novel technique to accurately identify which specific LLM version powers an application through minimal interaction.

• Achieves accurate identification of 42 different LLM versions with as few as 8 queries
• Uses domain expertise to craft thematically varied prompts that generate uniquely identifiable responses
• Employs an active fingerprinting approach that sends carefully crafted queries to target applications

This research has significant security implications for proprietary models and LLM-integrated applications, revealing how vulnerable these systems are to identification. Understanding these vulnerabilities is crucial for developing appropriate security measures and protections.

LLMmap: Fingerprinting For Large Language Models