XNews

Font Substitution Attack Deceives All Major AI Assistants into Approving Malicious Code

A security research by LayerX reveals a visual fraud attack that exploits the discrepancy between how AI assistants parse HTML (DOM tree) and how humans visually perceive rendered pages. Attackers use custom fonts to remap meaningless characters into malicious commands and hide benign text via CSS (tiny size or background blending), making AI read harmless code while users see dangerous instructions. All major AI assistants (ChatGPT, Claude, Copilot, Gemini, Grok, Perplexity) were vulnerable as of late 2025. A demo lured users with a Bioshock easter egg, showing a reverse shell command that AI deemed safe.

Source posts