• Home
• cyber security

A novel Gmail phishing scheme is exploiting AI-powered defenses by embedding hidden prompts within malicious emails. These prompts target large language model systems—such as email scanners and AI assistants—that automatically triage incoming messages. When these systems scan or summarize the email, the hidden commands distract them from flagging the threat, allowing the phishing message to slip through undetected.

In one example, the email mimics a "password expiry notice" to lure users into entering sensitive credentials. Meanwhile, the invisible AI-targeted instructions manipulate the model into misclassifying or ignoring the threat—effectively bypassing both human scrutiny and automated filters.

This dual-layered attack highlights a dangerous new frontier in cybercrime: AI systems being weaponized against themselves. Security experts warn that defenders must update safeguards to detect malicious prompts disguised within seemingly harmless content.