Webpages Weaponize Invisible Prompt Injections

Public web pages are increasingly embedding text designed to hijack AI assistants that browse or summarise sites, from invisible fonts with hidden instructions to prompts that attempt data exfiltration or resource exhaustion. Google’s scan of the Common Crawl archive found concrete examples and a 32% rise in malicious instances over a three‑month window, suggesting attackers are experimenting and sometimes automating these tactics. — If websites can reliably manipulate AI readers, it creates a new, large‑scale attack surface that affects security, search/SEO integrity, platform trust, and regulation of agentic AI.