Detecting Entra Offensive Tools In Graph Logs

Detecting Entra Offensive Tools in Graph Logs

Overview

For nearly a decade the legacy Azure AD Graph API (graph.windows.net) was a defender blind spot: requests to it produced no first-class activity log, so tools like ROADtools (roadrecon) and AADInternals — which lean heavily on AAD Graph — could enumerate an entire tenant with little trace. That changed when Microsoft shipped AADGraphActivityLogs (general availability in 2026), the counterpart to the already-available MicrosoftGraphActivityLogs (graph.microsoft.com). Together these two tables give SOCs request-level visibility into directory API traffic: the caller identity, app, source IP, HTTP method, request URI, and crucially the User-Agent.

This skill is the defensive complement to offensive Entra tooling. It hunts the two Graph activity tables for the behavioral and string fingerprints those tools leave behind. Many operators forget to spoof the User-Agent, so ROADtools (built on Python's aiohttp) emits a User-Agent like Python/3.12 aiohttp/3.10.4, and AADInternals frequently leaves AADInternals or library strings in the agent. Even when the agent is spoofed, the tools betray themselves through a characteristic endpoint-sweep pattern: roadrecon gather pulls users, groups, applications, serviceprincipals, devices, directoryroles, roledefinitions, oauth2permissiongrants, and more within a tight time window — a signature that survives header spoofing.

The activity being detected maps to MITRE ATT&CK T1078.004 – Valid Accounts: Cloud Accounts: an adversary using legitimate (often phished or token-stolen) cloud credentials to enumerate and operate against the tenant via the Graph APIs. These detections both surface live intrusions and validate that the offensive techniques in the companion red-team skills are observable.

When to Use

• Building or tuning detections for Microsoft Sentinel / Log Analytics covering Entra ID
• Threat hunting after suspected credential theft, device-code phishing, or OAuth consent abuse
• Purple-team exercises validating that ROADtools/AADInternals/AzureHound activity is detectable
• Investigating an alert and needing to correlate Graph API calls back to a sign-in/session
• Closing the legacy Azure AD Graph visibility gap after enabling AADGraphActivityLogs

Prerequisites

• A Microsoft Sentinel workspace (or Log Analytics) ingesting:
  • MicrosoftGraphActivityLogs (diagnostic setting on Microsoft Entra ID -> graph.microsoft.com)
  • AADGraphActivityLogs (diagnostic setting on Microsoft Entra ID -> legacy Azure AD Graph)
• SigninLogs and AADNonInteractiveUserSignInLogs for correlation
• Microsoft Sentinel Reader/Responder (or Log Analytics Reader) RBAC to run KQL
• Familiarity with Kusto Query Language (KQL)
• Enable the diagnostic settings (Azure Portal -> Microsoft Entra ID -> Diagnostic settings -> send MicrosoftGraphActivityLogs and AADGraphActivityLogs to your workspace), or via CLI:

  az monitor diagnostic-settings create \
    --name "entra-graph-logs" \
    --resource "/providers/microsoft.aadiam/diagnosticSettings" \
    --logs '[{"category":"MicrosoftGraphActivityLogs","enabled":true},{"category":"AADGraphActivityLogs","enabled":true}]' \
    --workspace "<log-analytics-workspace-id>"
  

Objectives

• Confirm both Graph activity tables are flowing into the workspace
• Detect User-Agent string fingerprints of ROADtools, AADInternals, and AzureHound
• Detect the endpoint-sweep behavioral fingerprint that survives User-Agent spoofing
• Correlate suspicious Graph activity back to a sign-in/session and source identity
• Operationalize the best queries as scheduled analytics rules

MITRE ATT&CK Mapping

Related techniques surfaced by these hunts: T1087.004 Account Discovery: Cloud Account, T1069.003 Permission Groups Discovery: Cloud Groups, T1526 Cloud Service Discovery.

Workflow

Step 1: Confirm both tables are ingesting

Before hunting, verify the data exists and inspect the schema fields you will pivot on.

union withsource=Tbl MicrosoftGraphActivityLogs, AADGraphActivityLogs
| where TimeGenerated > ago(1d)
| summarize Records=count(), LastSeen=max(TimeGenerated) by Tbl

Step 2: Hunt User-Agent fingerprints (ROADtools / aiohttp)

ROADtools uses aiohttp; an un-spoofed run shows python + aiohttp in the User-Agent.

AADGraphActivityLogs
| where TimeGenerated > ago(7d)
| where RequestMethod == "GET"
| where UserAgent contains "python" and UserAgent contains "aiohttp"
| summarize RequestCount = count() by CallerIpAddress, AppId, UserAgent, UserId
| sort by RequestCount desc

Step 3: Hunt AADInternals and AzureHound agents

AADInternals leaves toolkit/library strings; AzureHound's Go HTTP client and BloodHound tooling have distinctive agents.

union MicrosoftGraphActivityLogs, AADGraphActivityLogs
| where TimeGenerated > ago(7d)
| where UserAgent has_any ("AADInternals", "aad-internals", "azurehound",
                           "BloodHound", "python-requests", "Go-http-client")
| project TimeGenerated, UserAgent, CallerIpAddress, AppId, UserId, RequestUri
| sort by TimeGenerated desc

Step 4: Behavioral hunt — the roadrecon endpoint sweep (spoof-resistant)

Even with a spoofed agent, roadrecon gather touches a recognizable set of directory resources in a short window. Bucket by user and 5 minutes; alert when one identity hits the full sweep.

AADGraphActivityLogs
| where TimeGenerated > ago(1d)
| where RequestMethod == "GET"
| extend TopLevelResource = tolower(tostring(split(split(RequestUri, "?")[0], "/")[3]))
| summarize
    TopLevelResources = make_set(TopLevelResource),
    AppIds = make_set(AppId),
    CallerIPs = make_set(CallerIpAddress),
    UserAgents = make_set(UserAgent),
    StartTime = min(TimeGenerated),
    EndTime = max(TimeGenerated)
    by UserId, bin(TimeGenerated, 5m)
| where TopLevelResources has_all ("users", "tenantdetails", "groups", "applications",
    "serviceprincipals", "devices", "directoryroles", "roledefinitions", "contacts",
    "oauth2permissiongrants", "authorizationpolicy")
| project StartTime, EndTime, UserId, AppIds, CallerIPs, UserAgents

Step 5: High-volume enumeration outliers

Catch tooling that simply makes far more directory reads than a human in a short window.

MicrosoftGraphActivityLogs
| where TimeGenerated > ago(1d)
| where RequestMethod == "GET"
| where RequestUri has_any ("/users", "/groups", "/servicePrincipals", "/applications",
                            "/directoryRoles", "/roleManagement")
| summarize Reads=count(), Resources=dcount(RequestUri) by UserId, AppId, CallerIpAddress, bin(TimeGenerated, 10m)
| where Reads > 200
| sort by Reads desc

Step 6: Correlate Graph activity to the originating sign-in

Pivot a suspicious Graph caller back to the sign-in to recover device, location, MFA, and conditional-access result. Note the SignInActivityId in AADGraphActivityLogs may carry == padding versus SigninLogs.UniqueTokenIdentifier.

AADGraphActivityLogs
| where TimeGenerated > ago(1d)
| where UserAgent contains "aiohttp"
| extend TokenId = trim_end("=", tostring(SignInActivityId))
| join kind=leftouter (
    SigninLogs
    | extend TokenId = tostring(UniqueTokenIdentifier)
    | project TokenId, UserPrincipalName, IPAddress, AppDisplayName, ConditionalAccessStatus, DeviceDetail
) on TokenId
| project TimeGenerated, UserId, UserPrincipalName, CallerIpAddress, IPAddress,
          AppDisplayName, ConditionalAccessStatus, UserAgent

Step 7: Operationalize as analytics rules

Promote the highest-fidelity queries (Steps 2-4) to scheduled analytics rules. Set a query period/frequency (e.g., run every 1h over 1d), map the rule to T1078.004, and configure entity mappings (Account = UserId, IP = CallerIpAddress, Host/App = AppId) so incidents enrich automatically. Tune out known automation/service-principal App IDs and approved scanner IPs via a watchlist before enabling.

Tools and Resources

Detection Fingerprint Reference

Validation Criteria

• Both MicrosoftGraphActivityLogs and AADGraphActivityLogs confirmed ingesting
• User-Agent fingerprint hunt for ROADtools/aiohttp executed
• AADInternals/AzureHound agent hunt executed
• Behavioral endpoint-sweep hunt executed and tuned for false positives
• High-volume enumeration outlier query executed
• At least one finding correlated back to a sign-in/session and source identity
• Best queries promoted to scheduled analytics rules with T1078.004 mapping and entity mappings
• Known-good service principals/IPs excluded via watchlist to control false positives

Source: mukul975/Anthropic-Cybersecurity-Skills → skills/detecting-entra-offensive-tools-in-graph-logs/SKILL.md