Using LLM preferences for private key generation and model fingerprinting

Researchers demonstrate that copies of large language models can generate shared secret private keys or perform model fingerprinting over public channels by measuring how similarly they answer random binary preference questions. Testing models including Opus 5.5, Sonnet 5.5, GPT-6 Sol, and GPT-6 Luna, the author finds that identical model instances exhibit high preference correlation—largely driven by divisive questions—allowing them to establish secure information rates without exchanging obvious cryptographic data. Open article →