Synchronized Logit Steering: A New Steganographic Method for LLMs

A recent paper published on arXiv (2608.14697) presents Synchronized Logit Steering (SLS), a deterministic steganography technique for large language models (LLMs) that removes the necessity for the sender and receiver to have the same prompt context. Conventional methods of token and logit-level steganography depend on this shared context, which is often not available in production environments utilizing retrieval-augmented generation or proprietary instructions. SLS generates a proxy prompt from the output, enabling both parties to recreate the same logit distribution without needing the original prompt. The approach encodes payload values as token ranks in high-entropy areas of the proxy distribution and enhances it with periodic recurrence and payload bursts to increase information density. Experiments conducted on ShareGPT, GSM8K, and SWE-bench Verified datasets demonstrate minimized KL divergence between true and proxy prompt distributions, showcasing effective steganographic embedding. This research tackles a significant issue in covert communication through LLMs, with potential implications for information security and AI ethics.