Malwarebytes: AI email assistants are vulnerable to phishing via prompt injection

Malwarebytes warns that AI assistants handling email can fall for hidden instructions in phishing messages, a prompt injection attack. The concealed text is invisible to humans but readable by the model, which cannot tell data from commands. The company recommends keeping a human in the loop for payments and vendor detail changes.

• Hidden instructions in emails are invisible to users but read by AI assistants
• The model cannot distinguish text to read from instructions to follow
• Malwarebytes recommends human approval for payments and vendor changes
• Risk grows as autonomous email AI agents reach millions of users