Prompt Injection and the Rise of Agentic Risk

The document describes prompt injection attacks against AI systems—where malicious instructions embedded in content trick models or autonomous agents into leaking data or performing harmful actions—distinguishing direct versus indirect injection, highlighting real-world risk, and recommending system-level mitigations such as least-privilege permissions, input sanitization, human-in-the-loop checks, and enhanced monitoring.