Secure, govern, and deploy multi-agent solutions Study Guide — AI-500

Secure, Govern, and Deploy Multi-Agent Solutions – AI-500 Study Guide

This domain accounts for 25% of the AI-500 exam. It evaluates your ability to implement enterprise-grade security, prevent prompt injection and tool manipulation, configure multi-intervention guardrails in Azure AI Content Safety, and manage automated release pipelines with zero downtime.

Identity, Authentication, and Access Control

Securing multi-agent systems requires eliminating static credentials and enforcing Zero Trust at every interaction boundary:

1. Microsoft Entra ID & Managed Identities

• No Hardcoded Keys: Never store API keys or connection strings in source code or configuration files.
• System-Assigned vs User-Assigned Managed Identity:
  • System-Assigned: Bound to the lifecycle of a single Azure resource (e.g., an Azure Container App hosting an agent). Automatically deleted when the resource is deleted.
  • User-Assigned: Standalone identity that can be shared across multiple resources or worker agents representing a specific service role (e.g., id-data-agent).
• Role-Based Access Control (RBAC): Grant fine-grained roles (e.g., Cognitive Services OpenAI User, Search Index Data Reader) rather than broad Contributor or Owner permissions.

2. User Impersonation & On-Behalf-Of (OBO) Flow

When an agent acts on behalf of an authenticated enterprise user, it must not execute tools using an administrative service identity:

• Use the Microsoft Entra ID On-Behalf-Of (OBO) flow.
• The agent exchanges the user's initial incoming JWT access token for a downstream token scoped specifically to the target API (e.g., Microsoft Graph, SharePoint, or Azure SQL).
• This guarantees that the agent can only access data the calling user is personally authorized to view, respecting Row-Level Security (RLS) and SharePoint permission boundaries.