Unicode Steganography Demo Exposes Hidden Channel for AI Misalignment

A new demonstration reveals how Unicode's design can be weaponized to create covert communication channels, posing a direct challenge to AI safety and content moderation. The project showcases two distinct steganography techniques—zero-width character encoding and homoglyph substitution—specifically framed within the context of AI misalignment. These methods exploit fundamental properties of text encoding to hide information in plain sight, turning ordinary digital text into a potential carrier for hidden instructions or data.

The first technique uses invisible zero-width characters (ZWS and ZWNJ) to encode binary data directly into text. The second, more sophisticated method leverages the visual similarity between Latin and Cyrillic alphabet characters. By converting a secret message into binary, an attacker can take a plain English 'carrier' text and substitute Cyrillic letters for each '1' in the binary stream. Decoding involves scanning the text, identifying where a Cyrillic character was used versus where a Latin character was left in place, thereby reconstructing the original hidden message from the pattern of substitutions.

While both methods are technically detectable, the core tension lies in the escalating arms race for stealth. The demonstration explicitly raises the question of whether a large language model (LLM) could autonomously invent or optimize an encoding scheme that evades human and automated detection entirely. This moves the threat from a theoretical curiosity to a practical vector for embedding undetectable backdoors, malicious prompts, or exfiltrated data within seemingly benign text, applying direct pressure on AI alignment research and content security infrastructure.