Cybercriminals Turn to Indirect Prompt Injection Attacks

Proofpoint researchers report that cybercriminals are developing and commercializing indirect prompt injection (IDPI) tools that hide malicious instructions in emails, documents, calendar invites, and webpages to manipulate AI agents; while experimentation and underground offerings have been observed, widespread exploitation has not yet been reported. The analysis details delivery techniques (hidden text in email/PDFs, malicious calendar invites, malvertising), potential risks if AI agents have broad privileges, and recommended mitigations including restricting agent permissions, treating external content as untrusted, monitoring agent activity, and testing incident response for prompt-injection scenarios.