Security and Access Management

Quick Answer: Secure Azure AI solutions using managed identities for authentication, Key Vault for secret storage, RBAC for permission control, and private endpoints for network isolation. Microsoft Entra ID is preferred over API keys for production deployments.

Authentication Methods

API Keys (Simplest, Least Secure)

• Two keys provided per resource
• Passed in the Ocp-Apim-Subscription-Key header
• Suitable for development and testing only
• Risk: Keys can be leaked in source code, logs, or configuration files

Microsoft Entra ID / OAuth 2.0 (Recommended for Production)

• Token-based authentication using Azure identity
• Supports managed identities, service principals, and user identities
• No secrets to manage — tokens are obtained automatically
• Integrates with RBAC for fine-grained access control

Managed Identities (Best Practice)

• System-assigned: Created automatically with the Azure resource, tied to its lifecycle
• User-assigned: Created independently, can be shared across multiple resources
• Eliminates all credential management — Azure handles token issuance and rotation

# Using managed identity — NO keys or secrets in code
from azure.identity import DefaultAzureCredential
from azure.ai.textanalytics import TextAnalyticsClient

credential = DefaultAzureCredential()
client = TextAnalyticsClient(
    endpoint="https://my-language.cognitiveservices.azure.com/",
    credential=credential
)

On the Exam: When a question describes a production scenario requiring secure authentication without storing secrets, the answer is almost always managed identity with DefaultAzureCredential. API keys are acceptable only for development/testing scenarios.

Azure Key Vault Integration

Azure Key Vault provides centralized, secure storage for secrets:

Storing AI Service Keys in Key Vault

# Store an AI service key in Key Vault
az keyvault secret set \
    --vault-name my-keyvault \
    --name "AiServiceKey" \
    --value "a1b2c3d4e5f6..."

# Retrieve the key at runtime
az keyvault secret show \
    --vault-name my-keyvault \
    --name "AiServiceKey" \
    --query "value" -o tsv

Application Pattern

from azure.identity import DefaultAzureCredential
from azure.keyvault.secrets import SecretClient

# Retrieve AI service key from Key Vault
credential = DefaultAzureCredential()
secret_client = SecretClient(
    vault_url="https://my-keyvault.vault.azure.net/",
    credential=credential
)
ai_key = secret_client.get_secret("AiServiceKey").value

Role-Based Access Control (RBAC)

RBAC controls what actions users and applications can perform on Azure AI resources:

Assigning RBAC Roles

# Assign "Cognitive Services User" to a managed identity
az role assignment create \
    --assignee <managed-identity-object-id> \
    --role "Cognitive Services User" \
    --scope /subscriptions/<sub-id>/resourceGroups/<rg>/providers/Microsoft.CognitiveServices/accounts/<resource-name>

Network Security

Virtual Network Integration

• Restrict AI service access to specific virtual networks
• Deny all traffic from the public internet
• Allow only trusted Azure services to access the resource

Private Endpoints

• Create a private IP address for the AI service within your virtual network
• Traffic flows over the Microsoft backbone network, not the public internet
• DNS resolution maps the service endpoint to the private IP

Firewall Rules

# Restrict access to a specific virtual network
az cognitiveservices account network-rule add \
    --resource-group rg-ai-prod \
    --name my-ai-services \
    --vnet-name my-vnet \
    --subnet my-subnet

# Deny public access
az cognitiveservices account update \
    --resource-group rg-ai-prod \
    --name my-ai-services \
    --public-network-access Disabled

On the Exam: Private endpoints + disabled public access is the most secure network configuration. Questions may present a scenario where the AI service must not be accessible from the internet — the answer is private endpoints combined with disabling public network access.

Diagnostic Logging and Monitoring

Azure Monitor Integration

• Enable diagnostic settings to capture API call logs, errors, and metrics
• Send logs to Log Analytics workspace, Storage Account, or Event Hub
• Create alerts for error rate spikes, latency thresholds, or quota exhaustion

Key Metrics to Monitor