Researchers have developed a new method for embedding hidden messages in large language models called Synchronized Logit Steering (SLS). Unlike previous methods, SLS doesn't require the sender and receiver to share an identical prompt context. Instead, it uses the generated output itself to derive a proxy prompt that both parties can use to reconstruct the same logit distribution. This method is claimed to be more practical and stealthy than existing steganographic schemes in large language models. --- Why it matters: This matters because it enables covert communication through large language models without requiring shared context, which could have significant implications for applications such as secure data transmission or clandestine information exchange. Source: https://arxiv.org/abs/2608.14697