Security Dashboard for AI Security requirements explained (complex)
The new Microsoft Security Dashboard for AI (Preview) provides an end-to-end inventory and security risk overview of the entire AI apps and agents security estate.
The Microsoft Security Dashboard for AI (Preview) URL is:
https://ai.security.microsoft.com/The portal integrates Microsoft Entra, Microsoft Purview, and Microsoft Defender security products for AI. To ingest security data into the portal, almost all security products require configuration.
This blog will describe all the Security recommendations and configurations for the Security Dashboard for AI portal.
1️⃣Microsoft Entra
2️⃣Microsoft Purview
3️⃣Microsoft Defender
Microsoft Entra | Prevent agent sprawl and unauthorized access
Agent sprawl is the uncontrolled growth of AI Agents which introduces Security risks to the environment like unauthorized access, this can be internal or external.
Microsoft Entra Agent ID provides different Security products to mitigate against these risks of AI Agents (only the first two products below are required, and used for now, for the Security Dashboard for AI portal)
1️⃣ID Protection for Agents ✅
2️⃣Conditional Access for Agents ✅
3️⃣Identity Governance for Agents (out-of-scope)
4️⃣Secure Access for Agents (out-of-scope)
Disclaimer these Entra Agent ID products only support modern Agents.
See my previous blog for more details of all Entra Agent ID Security products.
Investigate risky agents with Entra ID Protection
ID Protection for agents provides an overview of risky Agents and protection for Agent Identities.
ID Protection for Agents is in Preview and can be used with an Entra Premium P2 or Microsoft 365 E5 license, this might change when the product goes to GA
This setting is enabled by default if the required license is available.
Enable Entra Conditional Access for agents
Conditional Access for Agent ID provides two features for Agent Identities / Agent Users:
1️⃣Block risky agents ❌
2️⃣Allow (included) agents to access (assigned) resources (e.g. via custom security attributes) ✅
The example below is the setup of a Conditional Access policy to block Agents with 🟥 high risk level.
You can use this template to create the Block High Risk Agent conditional access policy, and you can use this PowerShell script to verify the Conditional Access policy.
Conditional Access for Agents is in Preview and can be used with an Entra Premium P2 or Microsoft 365 E5 license, this might change when the product goes to GA
Microsoft Graph [or internal] API
Each product verification in the Security Dashboard for AI is using a Microsoft Graph or internal API call.
For example the verification of Enable Entra Conditional Access for agents uses the following GET request [F12 -> Network -> Fetch/XHR].
https://ai.security.microsoft.com/apiproxy/msgraph/beta/identity/conditionalAccess/policies?count=true&select=id&filter=....✅Status: 200 OK (successful)
The request is filtering Conditional Access policies with these parameters:
state eq 'enabled'— only active policies- AND involve Agent ID Service Principals in any way (
includeAgentIdServicePrincipals,excludeAgentIdServicePrincipals,agentIdServicePrincipalFilter) - OR target All Agent ID Users (
includeUsers in ('AllAgentIdUsers'))
When all Microsoft Entra Security products (in-scope) are enabled and configured, the requirements get status ✅
The Microsoft Graph API calls are essential for the next steps, where the security recommendations are high-level and broadly defined.
Microsoft Purview | Prevent data leaks and oversharing
Prevent data leaks and data oversharing is an important topic for internal threats, conscious or unconscious.
Microsoft Purview provides insights in data oversharing via Agents, these insights requires the security recommendations below.
Enable Microsoft Purview Audit
Microsoft Purview Audit is essential and required for user and agent activity.
1️⃣Sign in to the Microsoft Purview
2️⃣Go to Audit
3️⃣Start recording user and admin activity
API call verification for Admin Audit Log.
✅Status: 200 OK (successful)
Turn on Data Loss Prevention (DLP)
The security recommendation from Microsoft Purview DLP is very broad.
1️⃣Navigate to Data Loss Prevention
2️⃣Create a new policy
3️⃣Configure policy settings; define what sensitive information to protect, where to apply the policy, and what actions to take when violations occur.
As bespoken before this Security recommendation is very broad so I decided to get insights from the API calls on the Security Dashboard for AI page to find the exact match.
The exact validation of the setting is the Purview DLP Policy below.
Identity eq 'Default DLP policy - Protect sensitive M365 Copilot interactions'This DLP policy is part of the DLP policy recommendation Safeguard sensitive data in Copilot interactions. This policy protect items with sensitive data during Copilot interactions to prevent data leakage and oversharing by restricting Microsoft 365 Copilot and Copilot Chat from using sensitive info in prompts.
Turn on Communication Compliance, Insider Risk Management, and Data Lifecycle Management
The security recommendation from Microsoft Purview Communication Compliance, Insider Risk Management, and Data Lifecycle Management:
1️⃣Access Microsoft Purview solutions
2️⃣Configure Communication Compliance; set up policies to monitor communications for inappropriate content, regulatory violations, and policy breaches.
3️⃣Enable Insider Risk Management and Data Lifecycle Management; configure insider risk policies and retention schedules to automatically manage data throughout its lifecycle.
Same as the previous Purview requirements, the Security recommendations are even more broadly described. Let’s use the Developer Mode [F12] again to find the required Purview policies via the API calls.
Communication Compliance
The Communication Compliance policy API call requires the policy DSPM for AI — Unethical behavior in AI apps
Identity eq 'DSPM for AI - Unethical behavior in AI apps'✅Status: 200 OK (successful)
The required policy can be found in the DSPM for AI recommendations or in the templates of
Insider Risk Management
The Insider Risk Management policy API call requires
PolicyCategory eq 'ApplicableToAI'✅Status: 200 OK (successful)
The Insider Risk Management policy must be AI-related. The policy Detect risky interactions in AI apps (preview) is part of the DSPM for AI recommendations.
Data Lifecycle Management (Retention)
The Data Lifecycle Management (Retention) policy API call requires
PolicyCategory eq 'ApplicableToAI'✅Status: 200 OK (successful)
For AI-related retention policies, an Azure pay-as-you-go subscription is required. Beware of extra unpredictive cost
We can create an AI-related Retention Policy applied to Microsoft Copilot experiences.
When all Purview Security products and settings are enabled and configured, the requirements get status ✅
Microsoft Defender | Address AI risk and vulnerabilities
Address AI risk and vulnerabilities is protecting the organization proactive via security posture management and reactive via threat protection.
Microsoft Defender provides different products to provide protection for AI (AI Foundry & Copilot Studio) Agents against external threats.
Enable Defender for AI Services
Microsoft Defender for Cloud provides a plan called Defender for AI services to protect Azure AI Foundry agents.
1️⃣Sign in to the Azure portal and select Microsoft Defender for Cloud
2️⃣Select Environment settings and select the relevant Azure subscription
3️⃣Enable the AI services plan (Cloud Workload Protection) and Save
The Defender for AI Services is part of the Threat Protection for Agent IDs.
Configure cloud apps discovery
Cloud App discovery provides visibility in Shadow AI.
✅GenAI Apps; 1st (Copilot) and 3p (e.g. ChatGPT, Claude, etc.) GenIA apps
✅AI - Model providers; e.g. Anthropic Claude API
✅AI - MCP Servers
Analyze risk and apply Governance policies, manual or automatic. See my blog on GenAI Governance and the second blog about granular access.
1️⃣Enable Microsoft Defender for Cloud Apps in Microsoft Defender for Endpoint.
2️⃣Enable Microsoft Defender for Endpoint in Microsoft Defender for Cloud Apps.
To verify the setting in the Defender for Cloud settings, see Automatic log upload in the Cloud discovery section.
Discovered application data is flowing in and we can see Generative AI apps in-use in the SaaS apps section.
At the time of writing there’s a bug in the status of the Configure cloud apps discovery, it’s always disabled.
Even when I check the API call return values I get status Connected ✅
1️⃣ istenantonboarded = onboarding + toggle state
"isTenantOnboarded": true,
"toggleStatus": "on",
"isRegionSupported": true2️⃣ lcnc_settings = connection state (data flowing)
connectionStatus: "Connected"
inventoryConnectionStatus: "Connected"
agentProtectionIsConsent: true
isConsent: trueThe Cloud App discovery is part of the Security Posture management for Agent IDs.
Enable Microsoft 365 Connector
The Microsoft 365 connector is required for insights in Microsoft 365 user and Copilot Studio agent activities.
1️⃣Sign in to the Microsoft 365 Defender portal
2️⃣Go to App connectors in Microsoft Defender for Cloud Apps settings
3️⃣Connect and Select the Microsoft 365 connector
The Microsoft 365 connector is often already connected to monitor Microsoft 365 user activities
The Microsoft 365 Connector is part of the Security Posture management for Agent IDs.
Enable App Governance
Microsoft Defender for Cloud Apps provides App Governance, a security posture security feature to detect (e.g. high privileged) and stop (disable) risky cloud applications registered on Microsoft Entra ID.
1️⃣Sign in to the Microsoft 365 Defender portal
2️⃣Navigate to App Governance
3️⃣Enable the feature
The App Governance feature is part of the Security Posture management for Agent IDs.
Enable Copilot Studio AI agents
Microsoft Defender for Cloud Apps provides security features to protect AI Agents against threats.
1️⃣AI Agent Inventory✅, provides insight in security posture for agents like Authentication via the AIAgentsInfo table in Advanced Hunting, see blog for more details on how to enable this feature.
2️⃣Real-time protection during agent runtime (out-of-scope) protects agent tool abuse via (in)direct prompt injection, see blog for more details
3️⃣Block Image and URLs (out-of-scope) provides security against Image and URL abuse via indirect prompt injection on the internet, see blog for more details
All Defender for Cloud Apps and Copilot Studio threat protection features are currently in Preview and require a Microsoft 365 E5 license. This may change at GA.
When all in-scope Defender Agent ID security products are enabled and configured, the requirement status becomes ✅.
I hope this blog provides deeper insight into the security product recommendations for the Security Dashboard for AI portal.
