Implement security for AI
SharePoint oversharing, Purview DSPM for AI, real-time protection for Copilot Studio agents, Entra Agent ID access and Conditional Access, blast radius, AI gateway in API Management, Defender for AI services, Foundry guardrails, the Data and AI security dashboard, and managing agents.
- 28 min
- 11 concepts
- Try it yourself
- 5-question check
Learn
AI changes two things for security teams. Copilots and agents can surface any data a user can reach, so oversharing becomes a real risk. And agents are a new kind of identity that acts on its own, so it needs access control, monitoring and threat protection like any user or workload.
Several of these features are new; the lesson describes them as documented on Microsoft Learn, and links each source.
- 1
Identify overexposure of data in SharePoint
Use SharePoint Advanced Management data access governance reports (sites shared with 'Everyone except external users', sharing links, sensitivity-labelled files shared broadly), site access reviews by site owners, Restricted Content Discovery to hide sites from Copilot and search, and Restricted Access Control to limit sites to specific groups. Purview DSPM's oversharing assessments find exposed sensitive data.
Remember: Copilot shows what users can already access; fix permissions first.
- 2
Identify risks related to Microsoft Copilot and AI apps by using Microsoft Purview Data Security Posture Management (DSPM)
DSPM for AI (part of Purview DSPM) discovers AI usage across Microsoft 365 Copilot, agents and third-party AI sites, shows sensitive data in prompts and responses, runs data risk assessments for oversharing, and offers one-click policies (DLP for AI sites, insider risk for risky AI usage, collection policies).
Remember: Sensitive data in prompts? DSPM for AI shows it.
- 3
Enable and configure real-time protection for Microsoft Copilot Studio agents
Connect Copilot Studio to Microsoft Defender (via Defender for Cloud Apps and a Power Platform admin setting) so agent tool invocations are checked at runtime. Suspicious actions, such as prompt injection trying to make an agent call a tool to exfiltrate data, are blocked before they execute and raise alerts in the Defender portal.
Remember: Runtime protection blocks unsafe agent actions as they happen.
- 4
Implement conditional access for Microsoft Entra Agent ID
Conditional Access evaluates agents when they get tokens. Target the token subject: agent identities (or their blueprint, covering all agents created from it) for app-only access, agent users for agents with their own user account, and normal users for on-behalf-of access. Example: block high-risk agents. Policies don't carry across access patterns, and Users assignments don't cover agent users.
Remember: Target blueprints to cover future agents automatically.
- 5
Analyze blast radius for security risks related to Entra Agent ID by using Defender XDR
Defender XDR's incident graph and blast radius (powered by Sentinel graph) show what a compromised agent identity could reach: data, apps, other identities and critical assets, so you can prioritise containment and remove excessive permissions.
Remember: What could this agent reach if compromised? Blast radius.
- 6
Manage Entra Agent ID access
Microsoft Entra Agent ID gives agents first-class identities (agent identity blueprints, agent identities and agent users), each with sponsors/owners. Govern them like people: least-privilege permissions, access packages in entitlement management with expiry, lifecycle management, ID Protection for agents, and sign-in and audit logs.
Remember: Every agent needs an accountable human sponsor.
- 7
Configure and deploy AI Gateway in Azure API Management for Microsoft Foundry
API Management's AI gateway fronts Foundry and Azure OpenAI endpoints with policies: token limit (per consumer quotas), token metric emission, load balancing and circuit breaking across backends, semantic caching, content safety checks, and managed identity authentication to the backend, so keys aren't handed to apps.
Remember: Token limits per app: APIM AI gateway.
- 8
Enable Defender for AI Service in Cloud Workload Protection in Defender for Cloud
Defender for AI services detects threats to Azure AI and Foundry workloads at runtime: jailbreak attempts, credential theft, data leakage, wallet abuse and suspicious access, using Azure AI Content Safety Prompt Shields and threat intelligence. Optionally include suspicious prompt evidence in alerts. Alerts flow to Defender XDR.
Remember: Jailbreak alerts: Defender for AI services.
- 9
Configure guardrails for agent security in Foundry
In Microsoft Foundry, apply guardrails (content filters and controls) to models and agents: harmful content categories, Prompt Shields for direct and indirect (document) prompt injection, protected material detection, and tool-call controls, with block or annotate actions.
Remember: Indirect prompt injection: Prompt Shields.
- 10
Monitor AI security by using the Data and AI security dashboard in Defender for Cloud
The Data and AI security dashboard in Defender for Cloud gives one view of data and AI resources: discovered AI workloads (AI security posture management), sensitive data exposure, attack paths to AI resources, recommendations and active alerts, to prioritise risks.
Remember: One view of AI posture and threats in Defender for Cloud.
- 11
Manage agents in Microsoft 365 admin center
The agent registry (Agents in the Microsoft 365 admin center, part of Agent 365) lists Microsoft, partner and custom agents with owners, usage and risk signals from Defender, Entra and Purview. Admins approve or reject published agents, control who can use them, block risky agents immediately and reassign orphaned agents.
Remember: Block a risky agent instantly from the registry.
Who protects what in AI security
| Concern | Tool |
|---|---|
| Oversharing to Copilot | SharePoint Advanced Management, Purview DSPM |
| Sensitive data in prompts | Purview DSPM for AI, DLP |
| Agent identity and access | Entra Agent ID, Conditional Access |
| Unsafe agent actions | Copilot Studio real-time protection (Defender) |
| Attacks on AI workloads | Defender for AI services |
| Prompt injection | Foundry guardrails, Prompt Shields |
| API abuse and cost | APIM AI gateway |
Try it
Which control addresses each AI risk?
This interactive part needs JavaScript.
In the portal
- DSPM for AIMicrosoft Purview portal: Solutions > DSPM for AI
- Agent registryMicrosoft 365 admin center: Agents > All agents
- Conditional Access for agentsEntra admin center: Entra ID > Conditional Access > Policies > New policy
- Defender for AI servicesAzure portal: Microsoft Defender for Cloud > Environment settings > your subscription > Defender plans
Key terms
Click a card to see the definition.
Check your understanding
5 questions. Score 80% or more and these skills are marked as confident in your SC-500 study plan.
What's the first step before rolling out Copilot safely?
Copilot surfaces what users can already access.
Which Conditional Access target covers all agents created from one template, including future ones?
Blueprint policies apply to all its agent identities.
Which detects jailbreak attempts against Azure AI deployments?
Runtime threat protection for AI workloads.
Which APIM policy limits tokens per consumer?
AI gateway token limit.
Where are agents approved, blocked and reassigned?
The registry is the agent control plane.
Sources
- Microsoft Learn: Overview of Microsoft Agent 365
- Microsoft Learn: Conditional Access for agents
- Microsoft Learn: What is Microsoft Entra Agent ID?
- Microsoft Learn: Data Security Posture Management for AI
- Microsoft Learn: AI gateway in Azure API Management
- Microsoft Learn: Overview of AI threat protection
- Microsoft Learn: Prompt Shields
Facts in this lesson are from Microsoft Learn. This site isn't affiliated with Microsoft.