Show HN: LLM Steganography to hide text inside another plausible text (github.com)

A new implementation named Calgacus has been introduced, enabling LLM-based steganography through a method that conceals a secret message within a plausible text. This innovative approach, based on the MLX protocol—developed by Norelli and Bronstein in 2025—utilizes a language model's next-token distribution to encode information without altering the surface appearance of the output. Users can encode and decode messages by sharing a configuration file that contains essential parameters like the model ID, cover prompt, and other settings, ensuring that both sender and receiver can access the original message accurately. This steganographic technique is significant for the AI/ML community as it provides a practical method for secure communication and data concealment using existing language models. By leveraging the rank-driven selection process—where the encoder selects tokens based on their rank in the secret and cover distributions—the Calgacus implementation achieves lossless information preservation and maintains the natural flow of the cover text. This capability opens avenues for further research into covert communication methods and enhances the application landscape of AI models in privacy-driven applications. The implementation currently runs on Apple Silicon and requires Python 3.11+, making it accessible for developers interested in exploring this form of data encoding.