LLM Steganography: Same-Length Text Hiding

In this AI Research Roundup episode, Alex discusses the paper: 'LLMs can hide text in other text of the same length(2510.20075v2)' This work presents a simple, efficient protocol that lets LLMs hide a meaningful message inside a different, coherent text of exactly the same length. The authors show even 8B-parameter open models can encode and decode long messages locally in seconds. This capability decouples text from authorial intent and could be abused—for example, to mask unfiltered outputs within seemingly safe responses. The paper raises urgent questions for safety, oversight, and trust in written communication. Paper URL: https://arxiv.org/pdf/2510.20075 #AI #MachineLearning #DeepLearning #LLM #Steganography #ContentSafety #Misinformation #Security