LLMs and Text-in-Text Steganography

This article discusses attempts to hide text within LLMs using techniques like phonological changes and unconventional formatting (e.g., white text on white backgrounds). The author explores the limitations of these methods, noting that even relatively small LLMs can easily detect and decode them. The piece also touches on historical efforts to mitigate TEMPEST attacks using specialized fonts and software, highlighting advancements in SDR technology and the evolving landscape of signal intelligence. Ultimately, the author expresses skepticism about the capabilities of current LLMs, referencing criticisms from figures like Daniel Stenberg.