Governance Policies Azure AI Machine Learning

Machine Learning

Governance policies for Machine Learning

Domain: azure-ai

Patterns

Name	Description
ML workspace with managed network and CMK	Secure ML workspace with network isolation, CMK encryption, and associated resources

Anti-Patterns

Description	Instead
Do not deploy ML workspace without associated Key Vault, Storage, and App Insights	Always provision the four required dependency resources before workspace creation
Do not use workspace access keys for programmatic access	Use managed identity and RBAC role assignments (AzureML Data Scientist)
Do not deploy compute with public IPs	Set enableNodePublicIp=false and use managed network isolation

References

• Azure Machine Learning security baseline
• Configure managed network isolation

---

Checks (3)

Check	Severity	Description
AZ-ML-001	Required	Deploy Azure Machine Learning workspace with managed identity, high business impact, and no public access
AZ-ML-002	Required	Deploy compute instances and clusters with managed identity and no public IP
AZ-ML-003	Recommended	Use managed online endpoints with managed identity for model serving

---

AZ-ML-001

Deploy Azure Machine Learning workspace with managed identity, high business impact, and no public access

Severity: Required
Rationale: ML workspaces handle sensitive training data and models; managed identity eliminates credential sprawl
Agents: terraform-agent, bicep-agent, cloud-architect

Targets

• Microsoft.MachineLearningServices/workspaces

Companion Resources

Resource	Name	Purpose
Microsoft.Network/privateEndpoints	pe-ml-workspace	Private endpoint for ML workspace to eliminate public network exposure
Microsoft.Network/privateDnsZones	privatelink.api.azureml.ms	Private DNS zone for ML workspace API endpoint
Microsoft.Network/privateDnsZones	privatelink.notebooks.azure.net	Private DNS zone for ML workspace notebook endpoint
Microsoft.Insights/diagnosticSettings	diag-ml-workspace	Diagnostic settings to route ML workspace activity logs to Log Analytics
Microsoft.Authorization/roleAssignments	AzureML Data Scientist / Compute Operator	RBAC role assignments for data scientists and compute operators

---

AZ-ML-002

Deploy compute instances and clusters with managed identity and no public IP

Severity: Required
Rationale: Compute resources with public IPs and no identity create attack surface and credential risk
Agents: terraform-agent, bicep-agent, cloud-architect

Targets

• Microsoft.MachineLearningServices/workspaces

---

AZ-ML-003

Use managed online endpoints with managed identity for model serving

Severity: Recommended
Rationale: Managed endpoints handle scaling, versioning, and traffic splitting; managed identity secures model access
Agents: terraform-agent, bicep-agent, cloud-architect, app-developer, csharp-developer, python-developer

Targets

• Microsoft.MachineLearningServices/workspaces

---