LLMs and Text-in-Text Steganography: Security Implications

This article analyzes the capability of Large Language Models (LLMs) to perform text-in-text steganography, referencing a recent academic paper. The author demonstrates that LLMs can easily decode phonologically altered text, challenging assumptions about tokenization barriers. Even small models with four billion parameters handle these obfuscations effectively. The discussion expands to broader security concerns, including TEMPEST emissions and electromagnetic security (EmSec). It highlights how affordable Software Defined Radios (SDRs) and improved software like GNU Radio have made eavesdropping on monitor emissions easier, rendering older countermeasures like soft Tempest fonts less effective. The author recommends tools such as 'Tempest for Eliza' to demonstrate device insecurity. Additionally, the piece critiques the nature of LLMs, describing them as sophisticated autocomplete systems lacking true intelligence or intent. This perspective aligns with growing skepticism in the tech community, citing figures like cURL developer Daniel Stenberg who oppose AI hype. The article serves as a commentary on the evolving landscape of digital privacy, steganography, and the realistic limitations of current AI technologies.