Hackers Use Hidden Website Instructions in New Attacks on AI Assistants

Forcepoint X-Labs reports active "Indirect Prompt Injection" (IPI) attacks where threat actors hide adversarial instructions in ordinary web pages (via tiny fonts, comments, CSS, metadata, etc.) that autonomous LLM agents ingest and execute. Researchers found ten real-world examples in April 2026 showing impacts including financial fraud, data wiping, API key leaks, DoS and traffic hijacking, and identified trigger phrases and specific malicious sites used to activate these traps.