Private LLM copies can agree on a secret key over a public channel (matthewritch.com)

Recent research reveals that private instances of the same large language model (LLM) can efficiently establish a shared secret key over public communication channels through agreement on random preference questions, even under one-way communication. By leveraging the statistical similarity in their answers to these arbitrary questions, identical models can facilitate secure key exchanges. This breakthrough is significant for the AI/ML community as it paves the way for enhanced cryptographic methods using LLMs, enabling safer communication between model copies or their users while minimizing the risk of outside interference. The study involved testing several LLMs—Opus 5.5, Sonnet 5.5, GPT-6 Sol, and GPT-6 Luna—with 150 randomly generated preference questions. The results indicated a high degree of intra-model agreement, which averages about 90% across models. This agreement not only allows for the generation of secure keys but also for model fingerprinting, enhancing identification and authentication capabilities for AI applications. The framework also highlights the potential of constructing deliberately divisive questions to increase the efficiency of key generation. Overall, this research demonstrates the practical application of LLM behaviors in cryptography, opening new avenues for secure information exchange in AI-driven systems.