04 — Purview Integration

Configure Microsoft Purview to govern your ARIA assets with sensitivity labels, DLP enforcement, and lineage tracking.

What you’ll configure

• Purview sensitivity labels mapped to ARIA governance tiers
• A custom entity type for AI assets in the Purview Data Map
• DLP policies that prevent under-classified agents from accessing sensitive data
• Lineage edges tracking which agents consume which skills and knowledge

Prerequisites

• Completed 02 — Marketplace Setup
• Microsoft 365 E5 license (for Purview compliance features)
• Purview Account provisioned by Terraform
• Global Admin or Compliance Admin role for label configuration

Step 1: Map sensitivity tiers to Purview labels

ARIA uses five sensitivity tiers. Each maps to a Purview sensitivity label:

In the Purview compliance portal (compliance.microsoft.com):

1. Navigate to Information protection → Labels
2. Create or verify that labels matching each tier exist
3. Note the label GUIDs — you’ll use these in governance overlays

Step 2: Register AI asset entity type in Data Map

Purview’s Data Map needs a custom entity type to represent ARIA assets. This extends the Apache Atlas type system that Purview uses internally.

# Create the custom type definition
az rest --method POST \
  --uri "https://purview-aria-dev.purview.azure.com/catalog/api/atlas/v2/types/typedefs" \
  --headers "Content-Type=application/json" \
  --body '{
    "entityDefs": [{
      "name": "oasf_ai_asset",
      "description": "OASF-classified AI asset registered through ARIA",
      "superTypes": ["DataSet"],
      "serviceType": "ARIA",
      "attributeDefs": [
        {
          "name": "sensitivity_tier",
          "typeName": "string",
          "description": "ARIA sensitivity tier (public, internal, confidential, highly_confidential, restricted)"
        },
        {
          "name": "oasf_version",
          "typeName": "string",
          "description": "OASF Record version"
        },
        {
          "name": "asset_type",
          "typeName": "string",
          "description": "ARIA entity type (agent, skill, instruction, knowledge, orchestration)"
        },
        {
          "name": "compliance_frameworks",
          "typeName": "string",
          "description": "Comma-separated list of compliance frameworks"
        },
        {
          "name": "dependency_sensitivity_ceiling",
          "typeName": "string",
          "description": "Maximum sensitivity tier this asset is allowed to depend on"
        }
      ]
    }]
  }'

Step 3: Register relationship types

Define the ARIA relationship types so lineage edges appear in the Data Map:

az rest --method POST \
  --uri "https://purview-aria-dev.purview.azure.com/catalog/api/atlas/v2/types/typedefs" \
  --headers "Content-Type=application/json" \
  --body '{
    "relationshipDefs": [
      {
        "name": "aria_invokes",
        "description": "Agent invokes a Skill",
        "relationshipCategory": "ASSOCIATION",
        "endDef1": { "type": "oasf_ai_asset", "name": "invoker", "isContainer": false },
        "endDef2": { "type": "oasf_ai_asset", "name": "invoked_skill", "isContainer": false }
      },
      {
        "name": "aria_grounded_in",
        "description": "Agent is grounded in a Knowledge base",
        "relationshipCategory": "ASSOCIATION",
        "endDef1": { "type": "oasf_ai_asset", "name": "consumer", "isContainer": false },
        "endDef2": { "type": "oasf_ai_asset", "name": "knowledge_source", "isContainer": false }
      },
      {
        "name": "aria_governed_by",
        "description": "Agent is governed by an Instruction set",
        "relationshipCategory": "ASSOCIATION",
        "endDef1": { "type": "oasf_ai_asset", "name": "governed_agent", "isContainer": false },
        "endDef2": { "type": "oasf_ai_asset", "name": "instruction_set", "isContainer": false }
      }
    ]
  }'

Step 4: Configure DLP policies

Create a DLP policy that prevents under-classified agents from accessing data above their sensitivity ceiling. In the Purview compliance portal:

1. Navigate to Data loss prevention → Policies
2. Create a new policy scoped to the AI asset containers
3. Add a condition: “Content contains sensitivity label Confidential or above”
4. Add an action: “Block access when the requesting identity’s sensitivity_tier attribute is below confidential”

In practice, this is enforced through two mechanisms:

• CI/CD time: The oasf-validate workflow checks that the governance overlay’s sensitivity_tier doesn’t exceed its dependency_sensitivity_ceiling
• Runtime: The PurviewPolicyMiddleware in the Agent Framework evaluates prompts and responses against Purview DLP policies

Step 5: Update governance overlays with label IDs

Now that you have Purview labels, update your asset’s governance overlay with the actual label GUID:

{
  "governance": {
    "sensitivity_tier": "internal",
    "purview_label_id": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
    ...
  }
}

The purview-sync.yml workflow reads this label ID and applies it to the OCI artifact metadata in the Purview Data Map after each publish.

Step 6: Verify lineage in Purview

After publishing an asset, open the Purview governance portal:

1. Navigate to Data Map → Browse assets
2. Filter by type: oasf_ai_asset
3. Click on your asset to see its properties
4. Click the Lineage tab to see relationship edges

You should see your agent connected to its skills and knowledge bases via the aria_invokes and aria_grounded_in relationships.

Ongoing governance

Once configured, the Purview integration operates automatically:

• Every purview-sync.yml run registers/updates assets in the Data Map
• Sensitivity labels propagate through the dependency graph
• DLP policies evaluate AI interactions against declared classifications
• The Purview audit log captures all lifecycle events
• DSPM for AI surfaces anomalous usage patterns

The compliance team can now answer questions like:

• “Which agents access PHI-classified knowledge?”
• “If this knowledge base is reclassified, which agents are affected?”
• “Show me all assets governed under HIPAA with their approval chains”

Next Steps

→ 05 — Sample Agent