KQLHunting query: tune before alertingCloud AI abuse: Azure OpenAI, Vertex AI, self-hosted
Azure OpenAI model deployment or Responsible AI policy changed
Storm-2139 altered the capabilities of compromised generative AI services and bypassed safeguards. On Azure this maps to new or modified model deployments (accounts/deployments/write) and custom Responsible AI (content filter) policies being written or deleted (accounts/raiPolicies/write, raiPolicies/delete). The hunt flags callers who have not made such changes in the previous 30 days.
- Log source
- Microsoft Sentinel / Log Analytics: AzureActivity
- Techniques
- LLM10 Unbounded Consumption · AML.T0034 Cost Harvesting · AML.T0054 LLM Jailbreak
- MITRE ATT&CK
- T1685Disable or Modify ToolsT1496.004Cloud Service Hijacking
- Reviewed
- 2026-11-10
Query
let ops = dynamic([
"Microsoft.CognitiveServices/accounts/deployments/write",
"Microsoft.CognitiveServices/accounts/raiPolicies/write",
"Microsoft.CognitiveServices/accounts/raiPolicies/delete"]);
let known_callers = AzureActivity
| where TimeGenerated between (ago(30d) .. ago(1d))
| where OperationNameValue in~ (ops)
| distinct Caller;
AzureActivity
| where TimeGenerated > ago(1d)
| where OperationNameValue in~ (ops)
| where ActivityStatusValue in~ ("Succeeded", "Success")
| extend NewCaller = Caller !in (known_callers)
| summarize Operations = make_set(OperationNameValue), Resources = make_set(_ResourceId, 50),
CallerIPs = make_set(CallerIpAddress, 20), Changes = count()
by Caller, SubscriptionId, NewCaller
| order by NewCaller desc, Changes desc Expected false positives
- Planned model rollouts and content-filter tuning by the AI platform team
- Infrastructure-as-code pipelines applying routine changes
Sources
- Azure docs: Microsoft.CognitiveServices resource provider operations
- Azure Monitor: AzureActivity table reference
- Microsoft: Disrupting a global cybercrime network abusing generative AI (Storm-2139, Feb 2025)
Reviewed against vendor documentation, not tested in your environment. Defensive use only.