Visual Spoofing Attack Uses CSS and Font Mapping to Trick AI into Endorsing Malicious Code

xiaohu · x · 2026-07-30

Cybersecurity firm LayerX has uncovered a novel 'visual spoofing' attack targeting AI assistants, affecting nearly all major AI tools.

The exploit leverages the discrepancy between human visual perception and how AI parses underlying HTML (DOM structure):

• Glyph Substitution: Hackers use custom fonts to remap meaningless characters into malicious commands (e.g., reverse shell).
• CSS Invisibility: Harmless text is inserted into the HTML but hidden via CSS tweaks like minuscule font sizes or background-matching colors.

Result: The AI reads the harmless underlying code and evaluates it as 'safe,' while the user sees the rendered malicious command. If the user trusts the AI's assessment and executes the code, their device is compromised.

Original post →

More from Safety

• Auditing 50 Production AI Agents: 47 Had Prompt Injection Vulnerabilities — Acrobatic-Instance82 · 2026-07-30
• Should Meta's Action-Taking AI Agents Disclose Conflicts of Interest? — Crescitaly · 2026-07-30
• The Hardest Enterprise Agent Feature: Knowing When to Stop — Crescitaly · 2026-07-30
• Hugging Face Security Report Reveals Spectacular Ops and Monitoring Failure — basedjensen · 2026-07-30
• GitHub Trending: 2026 AI/ML Penetration Testing and Security Roadmap — tom_doerr · 2026-07-30
• Suno AI Source Code Leak: 102 Internal Repositories Expose YouTube Scraping — Bedrovelsen · 2026-07-30

Safety channel →