Cybersecurity experts warn criminals are tricking AI assistants with prompt injection

Cybersecurity experts warn that criminals are tricking AI assistants through prompt injection attacks, hiding malicious instructions in ordinary emails, webpages, or documents. Unlike traditional viruses, these attacks need no malware and can be written in plain English. AI agents may mistake hidden instructions for commands to follow, potentially revealing personal information or taking unauthorized actions. Experts say the threat exists now.

• OWASP: excessive agent authority is now a top LLM risk2026-09-16