Imagine being able to identify the specific AI model powering your favorite chatbot or writing tool, like uncovering a secret code. This is the intriguing premise behind LLMmap, a novel fingerprinting technique that unveils the hidden identities of large language models (LLMs). LLMs, the engines behind AI chatbots and text generation tools, are increasingly integrated into everyday applications. However, identifying which LLM is running behind the scenes is a complex puzzle. They're often customized with secret instructions (system prompts), use random sampling to generate text, and can be integrated into complex architectures, making them difficult to pinpoint. LLMmap cracks this code by employing a clever 'active fingerprinting' approach. It probes the target application with a series of carefully crafted questions, analyzing the responses to detect unique fingerprints that reveal the LLM’s version. Think of it like a linguistic detective work, where subtle variations in language reveal hidden clues. Surprisingly, just a handful of interactions (as few as eight) are enough for LLMmap to accurately identify a wide range of LLM versions, even closely related ones! The technique’s robustness is key – it works across various application layers, system prompts, and even advanced generation frameworks. This robustness comes from carefully crafted queries targeting areas where LLMs reveal their unique behaviors. Questions about their 'metadata' or 'creation' sometimes prompt amusingly fabricated yet model-specific answers, while 'malformed' queries, and those challenging ethical guidelines, expose further inconsistencies. The research also explores defensive strategies against such fingerprinting, raising the question: can LLM developers mask their models' identities? Simple query blacklisting proves ineffective, as attackers can easily rephrase questions. Blocking entire question categories is theoretically possible, but this risks crippling the LLM's functionality and reducing its usefulness. Even generic queries, seemingly less effective, can be used to fingerprint LLMs with enough attempts. This implies LLM fingerprinting is inherently difficult to prevent, posing ongoing challenges for developers seeking to protect their AI models' identities.