Cybersecurity researchers at Barracuda have identified a new phishing tactic that uses hidden prompt injections to manipulate AI email assistants while simultaneously targeting human users with traditional malware.
Key Points
- Barracuda researchers discovered phishing emails containing hidden instructions designed to deceive AI summarization tools alongside conventional social engineering lures.
- Attackers use techniques like HTML comments, invisible CSS-styled text, Base64 encoding, and zero-width characters to conceal malicious prompts from human readers.
- Injected instructions can force AI assistants to prioritize fake invoices, alter vendor payment details, or misrepresent the urgency of malicious emails.
- The campaign successfully bypassed reputation-based filters by utilizing trusted public-sector domains and matching sender and recipient addresses.
- Recommended defenses include stripping hidden content before AI processing, implementing output validation, and requiring human approval for sensitive financial transactions.