LLMmap puts its finger on ML attacks

Researchers demonstrated LLMmap, a fingerprinting technique that can identify 42 LLM versions with ~95% accuracy using as few as eight interactions, enabling attackers to move from generic probes to model-specific exploits such as tailored jailbreaks, prompt-injection, model extraction, privacy attacks, and potential architectural exploits; the article outlines defensive mitigations (rate-limiting, logging, least-privilege on tool access) while noting practical limits to fully preventing fingerprinting.