ricmmartins/azure-sre-agent-skills/08-ai-foundry-posture

AI Foundry & OpenAI Posture Check

Purpose

Assess the security, reliability, and cost efficiency of Azure OpenAI and Microsoft Foundry deployments. Detects the most common anti-patterns that startups make when building AI-powered products — from exposed endpoints to runaway token costs.

Based on the Azure Well-Architected Framework for AI workloads and Microsoft Foundry operational best practices.

When to use this skill

• User asks "is our OpenAI deployment secure?"
• User asks about AI cost optimization or token consumption
• Review before going to production with an AI feature
• User asks about content filtering, model versions, or rate limiting
• Periodic AI workload health check

Pre-check

Confirm with the user:

• Which Azure OpenAI / Cognitive Services accounts to assess (or "all in subscription")
• Whether they use PTU (Provisioned Throughput) or Standard deployments
• Whether they have production AI workloads already live

Assessment procedure

Step 0: Discover AI resources

az cognitiveservices account list \
  --subscription <sub-id> \
  --query "[?kind=='OpenAI' || kind=='AIServices'].{name:name, kind:kind, rg:resourceGroup, location:location, sku:sku.name}" \
  -o table

If no results, try:

az cognitiveservices account list \
  --subscription <sub-id> \
  --query "[].{name:name, kind:kind, rg:resourceGroup, location:location}" \
  -o table

If no Cognitive Services accounts exist, report "No Azure OpenAI or AI Foundry resources found" and end assessment.

For each account found, run the following checks:

🔐 CATEGORY 1 — Security (Critical)

Check 1.1 — Managed Identity enabled (not API keys only)

az cognitiveservices account show \
  --name <account> --resource-group <rg> \
  --query "{identity:identity.type, disableLocalAuth:properties.disableLocalAuth}" \
  -o json

📖 https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/managed-identity

Check 1.2 — Network isolation (firewall or private endpoint)

az cognitiveservices account show \
  --name <account> --resource-group <rg> \
  --query "{publicAccess:properties.publicNetworkAccess, defaultAction:properties.networkAcls.defaultAction, privateEndpoints:properties.privateEndpointConnections}" \
  -o json

📖 https://learn.microsoft.com/en-us/azure/ai-services/cognitive-services-virtual-networks

Check 1.3 — Content filtering (RAI policies)

az rest --method get \
  --url "https://management.azure.com/subscriptions/<sub-id>/resourceGroups/<rg>/providers/Microsoft.CognitiveServices/accounts/<account>/raiPolicies?api-version=2024-10-01"

Then check deployments have a policy assigned:

az cognitiveservices account deployment list \
  --name <account> --resource-group <rg> \
  --query "[].{name:name, model:properties.model.name, raiPolicy:properties.raiPolicyName}" \
  -o table

📖 https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/content-filters

⚙️ CATEGORY 2 — Reliability & Operations

Check 2.1 — Model version pinned and not deprecated

az cognitiveservices account deployment list \
  --name <account> --resource-group <rg> \
  --query "[].{name:name, model:properties.model.name, version:properties.model.version}" \
  -o table

Cross-reference with available models:

az cognitiveservices account list-models \
  --name <account> --resource-group <rg> \
  --query "[].{model:model.name, version:model.version, lifecycle:model.lifecycleStatus}" \
  -o table

📖 https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/model-retirements

Check 2.2 — Diagnostic settings enabled

# Get resource ID
RESOURCE_ID=$(az cognitiveservices account show \
  --name <account> --resource-group <rg> --query "id" -o tsv)
 
az monitor diagnostic-settings list --resource $RESOURCE_ID -o json

📖 https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/monitoring

Check 2.3 — Resource Lock on production account

az lock list --resource-group <rg> \
  --query "[?contains(id, '<account>')].{name:name, level:level}" \
  -o table

📖 https://learn.microsoft.com/en-us/azure/azure-resource-manager/management/lock-resources

Check 2.4 — Multi-region resilience (failover capability)

Check if OpenAI accounts exist in more than one region (disaster recovery / latency optimization):

az cognitiveservices account list \
  --subscription <sub-id> \
  --query "[?kind=='OpenAI' || kind=='AIServices'].{name:name, location:location, rg:resourceGroup}" \
  -o table

Group by location and count unique regions:

📖 https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/business-continuity-disaster-recovery

Context from the field: Customers with single-region deployments face outages during regional capacity events. Multi-region + APIM load balancing (Check 4.1) is the recommended pattern for production AI workloads.

Check 2.5 — 429 throttling rate (last 24 hours)

Requires diagnostic settings sending logs to Log Analytics (Check 2.2). If diagnostics are not configured, skip and note dependency.

AzureDiagnostics
| where TimeGenerated > ago(24h)
| where ResourceProvider == "MICROSOFT.COGNITIVESERVICES"
| where Category == "RequestResponse"
| summarize 
    TotalRequests = count(),
    ThrottledRequests = countif(resultSignature_d == 429),
    ThrottleRate = round(100.0 * countif(resultSignature_d == 429) / count(), 2)
  by Resource, properties_s
| order by ThrottleRate desc

Use execute_kusto_query with the Log Analytics workspace connected to the OpenAI account's diagnostic settings.

📖 https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/quota

Context from the field: Customers assume PTU eliminates throttling — it does not. Bursty traffic and concurrency limits can cause 429s even on PTU. Check APIM retry/spillover pattern (Check 4.1) as mitigation.

💰 CATEGORY 3 — Cost & Efficiency

Check 3.1 — Rate limits configured per deployment (not max)

az cognitiveservices account deployment list \
  --name <account> --resource-group <rg> \
  --query "[].{name:name, model:properties.model.name, sku:sku.name, capacity:sku.capacity}" \
  -o table

📖 https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/quota

Check 3.2 — Model selection diversity (not GPT-4o for everything)

az cognitiveservices account deployment list \
  --name <account> --resource-group <rg> \
  --query "[].{name:name, model:properties.model.name}" \
  -o table

📖 https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/models

Check 3.3 — PTU utilization (if applicable)

Only check if PTU/Provisioned deployments exist:

az cognitiveservices account deployment list \
  --name <account> --resource-group <rg> \
  --query "[?sku.name=='ProvisionedManaged'].{name:name, capacity:sku.capacity}" \
  -o table

If PTU deployments exist, check utilization via metrics:

az rest --method get \
  --url "https://management.azure.com<resource-id>/providers/microsoft.insights/metrics?api-version=2019-07-01&metricnames=ProvisionedManagedUtilizationV2&timespan=P7D&interval=PT1H&aggregation=Average"

📖 https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/provisioned-throughput

Check 3.4 — Quota utilization (TPM used vs limit)

az cognitiveservices account list-usage \
  --name <account> --resource-group <rg> \
  -o json

Alternative if the above returns empty (common for newer deployments):

az rest --method get \
  --url "https://management.azure.com/subscriptions/<sub-id>/providers/Microsoft.CognitiveServices/locations/<location>/usages?api-version=2024-10-01"

📖 https://learn.microsoft.com/en-us/azure/ai-services/openai/quotas-limits

Context from the field: Approved TPM quotas sometimes don't appear in Foundry due to subscription/project mismatches or the new Global/Data Zone quota pooling model. If usage data looks incorrect, verify the quota was assigned to the correct subscription and region.

Check 3.5 — Token consumption trend (tokens per day per deployment)

Requires diagnostic settings sending logs to Log Analytics (Check 2.2). If diagnostics are not configured, skip and note dependency.