Detect Azure Machine Learning Incorrect Authorization Information Disclosure (CVE-2026-68791) in Microsoft Sentinel
Detects potential exploitation of CVE-2026-68791, a critical (CVSS 10.0) incorrect authorization flaw (CWE-863) in Azure Machine Learning that allows an unauthorized remote attacker to disclose information over a network. Detection focuses on anomalous access to Azure ML workspace resources, data-plane API calls returning data to unexpected principals, cross-tenant or unauthenticated access patterns, and bulk retrieval of model artifacts, datastores, and compute secrets without corresponding RBAC role assignments. Because the flaw is a server-side authorization bypass, detection relies on Azure Monitor / activity and diagnostic logs showing successful reads from principals lacking appropriate permissions.
MITRE ATT&CK
- Tactic
- Discovery Collection Credential Access
KQL Detection Query
// df00tech DF00-6882a24dd9 · https://df00tech.com/detections/CVE-2026-68791 · licence: https://df00tech.com/detections/licence
// Azure ML workspace data-plane reads by principals without a workspace RBAC assignment
let lookback = 7d;
let knownAssignees = AzureActivity
| where TimeGenerated > ago(30d)
| where OperationNameValue has "Microsoft.Authorization/roleAssignments/write"
| where _ResourceId has "/providers/Microsoft.MachineLearningServices/workspaces/"
| extend assignee = tostring(parse_json(Properties_d).requestbody)
| summarize by Caller;
AzureDiagnostics
| where TimeGenerated > ago(lookback)
| where ResourceProvider == "MICROSOFT.MACHINELEARNINGSERVICES"
| where OperationName has_any ("DataStore", "Model", "Dataset", "Workspace", "ListKeys", "ListSecrets", "Download", "Read")
| where ResultType == "Success" or ResultSignature startswith "200"
| extend caller = coalesce(column_ifexists("identity_claim_appid_g", ""), column_ifexists("CallerIPAddress", ""))
| where isnotempty(caller) and caller !in (knownAssignees)
| summarize count(), reads = make_set(OperationName, 20), ips = make_set(CallerIPAddress, 10) by caller, Resource, _ResourceId, bin(TimeGenerated, 1h)
| where count_ > 10
| sort by count_ desc Finds successful Azure ML data-plane read/list operations (datastores, models, datasets, keys) from callers that do not appear in recent workspace role-assignment events, indicating possible authorization-bypass data disclosure.
Data Sources
Required Tables
False Positives & Tuning
- Automation service principals or managed identities with workspace access granted outside the observed role-assignment window (e.g. inherited from a parent scope).
- Data scientists performing legitimate high-volume dataset or model downloads during training runs.
- CI/CD pipelines (Azure DevOps / GitHub Actions) enumerating ML artifacts for deployment.
Other platforms for CVE-2026-68791
Testing Methodology
Validate this detection against 3 adversary techniques from Atomic Red Team. Each test below lists the behaviour to exercise and the telemetry you should expect to see. Executable commands and cleanup steps are further down this page.
- Test 1 Enumerate Azure ML workspace datastores via REST
Expected signal: AzureDiagnostics event with OperationName containing 'DataStore' and a List/Read action from the executing principal.
- Test 2 Retrieve datastore secrets (ListKeys/ListSecrets)
Expected signal: AzureDiagnostics event with OperationName containing 'ListSecrets' and ResultType Success.
- Test 3 Bulk model artifact download
Expected signal: Multiple AzureDiagnostics 'Model' Download events within a short window from one principal.
Response Playbook
Triage
- Identify the calling principal (appId/objectId) and source IP from the alert, then check Azure AD sign-in and audit logs to determine whether the identity is a known managed identity, service principal, or external/unknown actor.
- Enumerate the actual Azure RBAC role assignments on the affected ML workspace and its scope chain; confirm whether the principal legitimately holds read permissions (Reader, AzureML Data Scientist, etc.) or accessed data without an entitlement.
- Review the specific operations performed (ListKeys, ListSecrets, datastore/model Download) and the resources touched to assess the sensitivity and volume of data potentially disclosed.
- Determine whether the access originated from an unexpected tenant, IP range, or geography inconsistent with normal workspace usage.
Containment
- Apply or tighten RBAC: remove any erroneous assignments and, if an unauthorized principal is confirmed, disable or delete the service principal / revoke its credentials.
- Rotate any secrets, datastore access keys, and storage account keys exposed via ListKeys/ListSecrets operations, and restrict workspace network access with Private Endpoint / IP allowlists.
- If exploitation is confirmed and patch status is unknown, place the workspace behind a managed VNet and block public data-plane access pending Microsoft remediation guidance.
Evidence Collection
- Export AzureDiagnostics and AzureActivity records for the affected workspace covering the attack window, preserving operationName, caller, callerIPAddress, resultType and resourceId fields.
- Capture Azure AD sign-in/audit logs for the implicated principal and snapshot the current RBAC role-assignment state of the workspace and parent scopes.
Escalation Criteria
- !Escalate to incident response if secrets or datastore keys were retrieved (ListKeys/ListSecrets success) or if the principal is confirmed unauthorized/external.
- !Escalate to the cloud security and data governance teams if sensitive training data, models, or customer datasets were downloaded, triggering breach-notification assessment.
Investigation Guide
Forensic Artifacts
- >
AzureDiagnostics entries for MICROSOFT.MACHINELEARNINGSERVICES with caller, operationName and resultType. - >
Azure AD sign-in logs and service-principal credential usage for the implicated principal. - >
Workspace RBAC role-assignment history and Activity Log roleAssignments events.
Tuning Guidance
Baseline the set of managed identities and service principals that routinely access each ML workspace and exclude them from the unexpected-principal logic. Adjust the reads>10 threshold to match legitimate training-job download volumes in your environment, and prioritise alerts involving ListKeys/ListSecrets or access from IPs outside your corporate/VNet ranges.
Hunting Queries
Hunts for successful secret/key enumeration against Azure ML workspaces, the highest-impact variant of the disclosure.
// df00tech DF00-6882a24dd9 · https://df00tech.com/detections/CVE-2026-68791 · licence: https://df00tech.com/detections/licence
AzureDiagnostics | where ResourceProvider == "MICROSOFT.MACHINELEARNINGSERVICES" | where OperationName has_any ("ListKeys","ListSecrets") | where ResultType == "Success" | summarize count() by CallerIPAddress, OperationName, _ResourceId, bin(TimeGenerated,1h) | sort by count_ desc ``` df00tech DF00-6882a24dd9 · https://df00tech.com/detections/CVE-2026-68791 · licence: https://df00tech.com/detections/licence ```
index=azure ResourceProvider="MICROSOFT.MACHINELEARNINGSERVICES" (operationName="*ListKeys*" OR operationName="*ListSecrets*") resultType=Success | stats count by callerIpAddress operationName resourceId Atomic Red Team Tests
Simulates reconnaissance of ML datastores using the management API; in a lab, run with and without appropriate RBAC to validate telemetry.
Command
az ml datastore list --workspace-name lab-mlws --resource-group lab-rg -o table Cleanup
echo 'No state change; read-only enumeration.' Expected Telemetry
AzureDiagnostics event with OperationName containing 'DataStore' and a List/Read action from the executing principal.
Expected Detection
KQL/SPL rule records the list operation; correlation fires if followed by downloads.
Attempts to list datastore credentials, the key disclosure behaviour of CVE-2026-68791.
Command
az rest --method post --url "https://management.azure.com/subscriptions/$SUB/resourceGroups/lab-rg/providers/Microsoft.MachineLearningServices/workspaces/lab-mlws/datastores/default/listSecrets?api-version=2023-10-01" Cleanup
echo 'Read-only; rotate any keys retrieved in production labs.' Expected Telemetry
AzureDiagnostics event with OperationName containing 'ListSecrets' and ResultType Success.
Expected Detection
Hunting query for ListKeys/ListSecrets fires on successful secret retrieval.
Downloads multiple registered models to simulate bulk data disclosure.
Command
for ($i=1; $i -le 15; $i++) { az ml model download --name testmodel$i --version 1 --workspace-name lab-mlws --resource-group lab-rg --download-path C:\temp\models } Cleanup
Remove-Item -Recurse -Force C:\temp\models Expected Telemetry
Multiple AzureDiagnostics 'Model' Download events within a short window from one principal.
Expected Detection
reads>10 threshold in the KQL/SPL/CQL rules triggers on the bulk-download volume.