Secure AI agents against prompt injection, tool abuse, and data exfiltration with defense-in-depth controls.