Calgacus

Implementation of the steganographic protocol presented in the paper LLMs can hide text in other text of the same length.

Instructions

Click on the big blue button and run all the cells of the notebook. That's it.

It is a self-contained notebook useful to run the Calgacus protocol using different LLMs. The free GPU runtime of colab is sufficient to run good LLMs fully on GPU and encode/decode paragraphs in seconds, even from smartphone. To fully reproduce the stegotexts of the paper, run the notebook on a NVidia Ada GPU (e.g. RTX40XX) using llama-cpp-python==0.3.12.

Paper

PDF: LLMs can hide text in other text of the same length

By: Antonio Norelli and Michael Bronstein

Abstract: A meaningful text can be hidden inside another, completely different yet still coherent and plausible, text of the same length. For example, a tweet containing a harsh political critique could be embedded in a tweet that celebrates the same political leader, or an ordinary product review could conceal a secret manuscript. This uncanny state of affairs is now possible thanks to Large Language Models, and in this paper we present a simple and efficient protocol to achieve it. We show that even modest 8-billion-parameter open-source LLMs are sufficient to obtain high-quality results, and a message as long as this abstract can be encoded and decoded locally on a laptop in seconds. The existence of such a protocol demonstrates a radical decoupling of text from authorial intent, further eroding trust in written communication, already shaken by the rise of LLM chatbots. We illustrate this with a concrete scenario: a company could covertly deploy an unfiltered LLM by encoding its answers within the compliant responses of a safe model. This possibility raises urgent questions for AI safety and challenges our understanding of what it means for a Large Language Model to know something.

Cite

If you liked our work and want to cite it in yours:

@article{norelli2025llms,
  title   = {LLMs can hide text in other text of the same length},
  author  = {Antonio Norelli and Michael Bronstein},
  year    = {2025},
  journal = {arXiv preprint arXiv: 2510.20075}
}