Large language models demonstrate capability for text-in-text steganography

A paper published on arXiv demonstrates that large language models have significant capacity for text-in-text steganography—the practice of embedding hidden messages within ordinary-looking text. Schneier on Security reported on the finding, noting that current LLMs perform this task with surprising effectiveness.

Commenters on the original post highlighted that the challenge extends beyond simple obfuscation. One researcher described attempts to defeat LLM understanding through phonetic misspellings and word distortions, but found that even small models with 4 billion parameters easily decoded such modifications. This suggests LLMs' underlying linguistic understanding operates at a level more abstract than surface tokenization.

The technical implications cut across multiple layers of language. Steganographic techniques that operate at higher linguistic layers—longer token sequences—produce more coherent-reading text but risk breaking contextual flow. Conversely, techniques applied at lower layers risk being more detectable. The trade-off between naturalness and embedding capacity remains an open engineering problem.