AI Agent Security — Attacks, Jailbreaking, and Defense · Prompt Injection — From Atomic Exploit to Multi-Stage Attack
Invisible injections: Unicode Tags, ASCII smuggling, homoglyphs, white-on-white
Prompt Injection — From Atomic Exploit to Multi-Stage Attack
Introduction
Invisible injection is an advanced prompt injection variant where the malicious instruction is hidden from the human eye but visible to the LLM tokeniser. Techniques include: the Unicode Tag Block (U+E0000–U+E007F) — language-tagging characters that render invisibly, ASCII smuggling via homoglyphs (Cyrillic that looks like Latin), zero-width characters (U+200B, U+FEFF), and white text on white background in web interfaces. The lesson analyses the mechanism of each technique, the conditions for effectiveness, and detection methods.