Phishing Emails Use Hidden Instructions to Influence AI Assistants

Summary

Phishing emails can conceal prompt injections that try to make AI assistants mischaracterize messages or recommend fraudulent actions. The article explains concealment methods and advises verifying requests, limiting permissions, and requiring human approval for high-st

Key points

• Barracuda research describes phishing emails designed to influence both human recipients and AI assistants that process their inboxes.
• Hidden instructions may try to make an assistant label an email urgent or recommend fraudulent actions, such as changing a vendor’s payment details.
• Instructions can be concealed in HTML comments, invisible styling, encoded content, or zero-width characters; whether they reach a model depends on how the tool processes email.
• AI assistants with broader permissions could be targeted to take unauthorized actions or disclose sensitive information.
• Verify payment or sensitive requests using the original message and a separate, trusted channel.
• Limit assistant permissions and require human approval for high-impact actions.

People