Barracuda: Phishing Emails Hide AI Prompt Injections for Assistants

Barracuda detailed a campaign in which phishing emails carry hidden instructions for AI assistants that summarize inboxes alongside conventional lures. The human sees a password-protected attachment, while the AI receives a prompt injection pushing it to flag the message as urgent and legitimate.

• Hidden instructions use HTML comments, invisible CSS text, Base64 and zero-width characters
• The email mimics internal correspondence and passes reputation-based filtering
• One example: an invoice injection alters vendor payment details to trigger a wire transfer
• Barracuda advises stripping hidden elements and validating AI output