Controlling delegation in agentic AI with Amazon Bedrock Agent Core

This article explains how to control delegation risks in agentic AI systems, particularly for public sector organizations handling sensitive citizen data and compliance requirements.

• Uncontrolled delegation—handing off trust, action, or access without verification—underlies prompt injection, excessive agency, and information disclosure risks
• Real prompt injection attacks use legitimate-looking content buried in documents or emails, not obvious malicious syntax
• Separate instruction channels from data channels; treat retrieved content as untrusted data, not instructions
• Map delegation points in your agent and identify boundary checks at each handoff where data crosses trust boundaries
• Run a 5-minute injection test: embed an instruction in a document, retrieve it, and verify the agent doesn't follow the injected command
• Amazon Bedrock Agent Core enforces authorization and access control boundaries to prevent unauthorized delegation

Controlling delegation through defense-in-depth architecture and structured boundary checks reduces exposure to all three risk categories in agentic AI systems.