Abstract <p>We propose a novel framework for steganographic text generation that hides binary messages within semantically coherent natural language using latent-space conditioning of large language models (LLMs). Secret messages are first encoded into continuous vectors via a learned binary-to-latent mapping, which is used to guide text generation through prefix tuning. Unlike prior token-level or syntactic steganography, our method avoids explicit word manipulation and instead operates entirely within the latent semantic space, enabling more fluent and less detectable outputs. On the receiver side, the latent representation is recovered from the generated text and decoded back into the original message. As a key theoretical contribution, we provide a robustness guarantee: if the recovered latent vector lies within a bounded distance of the original, exact message recovery is ensured, with the bound determined by the decoder’s Lipschitz continuity and the minimum logit margin. This formal result offers a principled view of the reliability–capacity trade-off in latent steganographic systems. Empirical evaluation on both synthetic data and real-world domains such as Amazon reviews shows that our method achieves high message recovery accuracy (above 91%), strong text fluency and competitive capacity up to 6 bits per sentence element while maintaining resilience against neural steganalysis. These findings demonstrate that latent conditioned generation offers a secure and practical pathway for embedding information in modern LLMs.</p>