New Steganography Method Hides Messages in LLM Text Without Shared Prompts

This research introduces Synchronized Logit Steering (SLS), a novel steganography technique designed for large language models. Unlike previous methods that rely on both parties having identical prompt contexts, SLS overcomes this limitation by deriving a proxy prompt directly from the generated text. This allows both the sender and receiver to reconstruct the necessary logit distribution independently, facilitating covert communication in real-world scenarios where prompt sharing is impractical. The SLS method encodes information by mapping payload values to token ranks within high-entropy regions of the proxy prompt distribution. The study demonstrates that this technique achieves robust synchronization after about 40 tokens and can reach a capacity of 0.20 bits per token with periodic-burst encoding, significantly higher than single-payload methods. Statistical tests confirm that SLS-generated outputs are nearly indistinguishable from standard greedy generations, highlighting its stealth and practicality for hidden communication.