Indirect Prompt Injection Attacks: A Lurking Risk to AI Systems

This report explains how indirect prompt injection has emerged as a significant TTP targeting GenAI systems by hiding adversarial instructions in external content (emails, documents, webpages, images) that AI tools ingest, enabling data exfiltration, process manipulation, and potential lateral movement. It highlights the expanded attack surface from shadow AI and provides a multilayer defense approach: prompt injection detection, input validation and sanitization, content security policies and allowlisting, least-privilege access and confirmation for risky actions, monitoring and governance of AI usage, and user education. The piece also notes CrowdStrike’s AI detection and response capability for mitigating these attacks.