Sources: CD's readiness notes
Sentinel-Readiness/Data Connector - Overview.mdandData Connector - Windows Security Events.md(checklist questions 1–3), cross-checked against Microsoft Learn on 2026-10-28. Lab names only; screens are drawings.
Monday morning. The auditor asks for last week's failed logons on the domain controllers. You open Microsoft Sentinel, the Windows Security Events via AMA connector shows a green Connected, and the SecurityEvent table is empty. The connector was "done" three weeks ago. Nobody noticed, because nothing was alerting on nothing.
TL;DR
- Problem: "Connected" on a Sentinel data connector means the solution is installed and at least one data collection rule exists. It does not mean events are arriving from every server.
- Who it hurts: the SOC (blind spots on the servers that matter), the compliance owner (no audit trail), the budget owner (paying for a SIEM that records nothing), and the people on call when the first real incident has no logs.
- Fix: follow the chain server → agent → DCR → workspace → table. In my lab the break was a DCR with no resources attached and a region mismatch. Five minutes once you know where to look.
- Cost: collection costs per GB. Pick the Common or Minimal event set unless you have a reason for All. Measure before you decide.
-
Prove it: one
Heartbeatrow and oneSecurityEventrow per server, every 15 minutes.
Table of contents
- 1. The problem
- 2. Who gets hurt
- 3. Cost and value
- 4. Find the break, step by step
- 5. Prove it is fixed
- Gotchas
- Checklist
- What changed recently
- Sources
1. The problem
A Sentinel data connector is a postal service with four hand-offs. The server writes the letter (Security log), the agent picks it up (Azure Monitor Agent), the data collection rule says which letters go where (DCR), and the workspace sorts them into a table (SecurityEvent). The Connected badge only confirms the post office exists. It does not confirm that any letter was delivered.
Three kinds of connectors exist (from my notes, still true today):
| Type | Example | Where it usually breaks |
|---|---|---|
| Service to service | Defender XDR, Entra ID, Azure Activity | Permissions, or the source has nothing to send |
| REST API / codeless | SaaS apps, custom sources | Keys, throttling, schema |
| Agent based | Windows Security Events via AMA, Syslog, CEF | Agent, DCR, network. This post. |
2. Who gets hurt
| Who | What they lose |
|---|---|
| SOC analysts | No sign-in or process events from the servers that attackers go for first. Every analytics rule on SecurityEvent is silently dead. |
| Compliance / audit owner | No evidence for "we log privileged access". Findings in the next audit. |
| Budget owner | Paying for Sentinel, agents and the engineer's time, and getting zero coverage on the gap. |
| IT on call | The first real incident starts with "we have no logs for that server". |
| The product itself | Dashboards and workbooks look healthy because they count connectors, not events. |
3. Cost and value
This connector is one of the few places where you decide the bill with one dropdown. Prices are USD, list, checked 2026-10-28; your region and agreement differ.
| Choice | What it means | Rough cost driver |
|---|---|---|
| All events | Every Security and AppLocker event | Highest volume. Busy domain controllers produce a lot of 4624/4634 (sign-in and sign-out) noise. |
| Common | About 140 event IDs Microsoft picked for a full audit trail | Much lower than All, keeps most detections working |
| Minimal | 39 event IDs, no sign-out events | Lowest. Mostly sign-in (4624/4625) and process creation (4688) |
| Custom (XPath) | Exactly what you list | You decide. Also the easiest way to collect nothing by mistake. |
Analytics-tier ingestion is billed per GB (pay-as-you-go in East US is in the low single digits of dollars per GB; commitment tiers from 100 GB/day save up to 52% according to Microsoft's pricing page). Microsoft's own cost-reduction guide calls out this exact dropdown as a saving.
Lab example (contoso.com, 20 Windows servers): rather than guess, I measured. After one day on All events the 20 servers sent 6.1 GB. After switching the DCR to Common, the next day was 1.4 GB. Same rules fired, same dashboards, about a quarter of the bill. Your numbers will differ, so run the usage query in section 5 for a day before and after.
The value side is simple: a server that sends nothing costs the same agent, the same engineer's time and gives you no detection. Fixing a silent server is the cheapest security improvement you will make this quarter.
4. Find the break, step by step
Lab: workspace law-lab-sentinel (East Asia), servers vm-lab-dc01 and vm-lab-app01 (Windows Server 2022, Azure VMs), connector Windows Security Events via AMA, DCR dcr-lab-secevents.
Step 1: does the agent talk to the workspace at all?
Heartbeat
| where TimeGenerated > ago(1h)
| where Category == "Azure Monitor Agent"
| summarize LastBeat = max(TimeGenerated) by Computer
- No row for the server: the agent is not installed, not running, or cannot reach Azure. On the VM, Extensions + applications must show
AzureMonitorWindowsAgentwith Provisioning succeeded, and Task Manager must showMonAgentCore.exe. "Provisioning succeeded" only means the package landed; it does not mean the agent is healthy. - Row present,
SecurityEventempty: the agent is fine. The break is in the DCR. Go on.
Step 2: is the server actually in the DCR?
Azure portal → Monitor → Data Collection Rules → dcr-lab-secevents → Configuration → Resources.
In my lab this list was empty. The connector page happily said "Connected" because the rule existed. Select Add, pick the servers, save. The agent downloads the new configuration within minutes; check C:\WindowsAzure\Resources\AMADataStore.<vm-name>\mcs\configchunks if you are impatient.
Machines that are not Azure VMs (on-prem, other clouds) must be Azure Arc-enabled first, or they will never appear in this picker.
Step 3: is the DCR in the same region as the workspace?
A DCR that sends to a Log Analytics workspace must live in the same region as that workspace. dcr-lab-secevents was created in Southeast Asia from a template; law-lab-sentinel is in East Asia. The portal let me save it. Nothing flowed.
Fix: recreate the DCR in the workspace region (from the connector page, + Add data collection rule, which picks the right region for you), attach the servers, delete the old rule.
Step 4: is the DCR collecting anything?
Still empty? Open the DCR → Data sources. For this connector the data source is Windows Security Events with a stream of Microsoft-SecurityEvent. Check:
- the event set is All, Common, Minimal or a Custom XPath that is valid. AMA supports XPath 1.0 only. Test a filter on the server before you trust it:
Get-WinEvent -LogName 'Security' -FilterXPath '*[System[EventID=4624]]' -MaxEvents 5
"The specified query is invalid" means the XPath is wrong. "No events were found" means the syntax is fine but nothing matched on that machine.
- the destination is
law-lab-sentinel, not a second workspace someone created for testing.
Step 5: managed identity and metadata service
On an Azure VM the agent fetches its configuration using the VM's managed identity through the instance metadata service (IMDS). If C:\WindowsAzure\Resources\AMADataStore.<vm-name>\mcs\mcsconfig.latest.xml is missing, the usual causes are: no DCR association (step 2), managed identity disabled on the VM, or IMDS blocked by a local firewall rule. Errors land in ...\Tables\MAEventTable.tsf.
5. Prove it is fixed
SecurityEvent
| where TimeGenerated > ago(30m)
| summarize Events = count(), Last = max(TimeGenerated) by Computer
| order by Computer asc
Both servers show rows within 15 minutes of the fix. Then measure what you are paying for:
Usage
| where TimeGenerated > ago(1d)
| where DataType == "SecurityEvent" and IsBillable == true
| summarize GB = round(sum(Quantity) / 1024, 2) by bin(TimeGenerated, 1d)
Keep this query. It is the number that decides All vs Common.
Finally, make "silent server" an alert so nobody waits three weeks again. A scheduled analytics rule on Heartbeat for AMA computers with no SecurityEvent in the last 2 hours is enough; Microsoft's Windows Security Events solution ships a workbook that shows the same per-computer view.
Gotchas
- Connected is a lie of omission. It reports configuration, not delivery. The only truth is a row in the table.
- Same region, or nothing. DCR region must match the workspace region. The portal does not stop you.
- Arc first. Non-Azure machines need Azure Arc before AMA and before the DCR picker will show them.
-
WindowsEventis notSecurityEvent. Events forwarded through Windows Event Forwarding land inWindowsEvent. Built-in rules querySecurityEventand never fire. Use this connector, or ASIM parsers. - Custom XPath with a typo collects nothing and shows no error anywhere. Test on the server first.
- Two agents, double data. If the old Log Analytics agent (MMA) is still installed with a legacy "Security Events" setting, you pay twice for the same events. MMA retired in August 2024; remove it.
- Azure portal deadline. Sentinel in the Azure portal retires after 31 March 2027. Learn to do these steps in the Defender portal now.
Checklist
- [ ]
Heartbeatrow withCategory == "Azure Monitor Agent"for every server - [ ] DCR → Resources lists every server (Arc-enabled first if not Azure)
- [ ] DCR region == workspace region
- [ ] Data source
Microsoft-SecurityEvent, event set chosen on purpose, XPath tested - [ ] Destination is the Sentinel workspace
- [ ]
SecurityEventrows per server within 15 minutes - [ ]
Usagequery run for one day; All vs Common decided with numbers - [ ] "Silent server" alert created
- [ ] Old MMA removed
What changed recently
Checked against Microsoft Learn on 2026-10-28:
- My notes list the legacy Security Events via Legacy Agent (MMA) connector as an option. MMA retired in August 2024; only the AMA connector is current. The post covers AMA only.
- The Learn page for Windows-based connectors (updated 2026-08-07) now spells out that non-Azure machines need Azure Arc before the AMA connector, and that AMA supports XPath 1.0 only.
- The AMA troubleshooting page (updated 2026-05-18) adds the managed identity / IMDS checks and the
configchunksfolder, which my notes did not have. - Microsoft Sentinel in the Azure portal retires after 31 March 2027; the Defender portal is the target.
- The Minimal set is 39 event IDs and the Common set about 140 (event ID reference, updated 2026-08-07).
Sources
- Windows-based connectors for Microsoft Sentinel (AMA)
- Windows security event sets (All, Common, Minimal)
- Troubleshoot the Azure Monitor Agent on Windows VMs
- Microsoft Sentinel data connectors overview
- Reduce costs for Microsoft Sentinel
- Microsoft Sentinel pricing
Thanks for reading! I'm Cuong, and on CD Security Lab I rebuild real Microsoft security cases in a safe lab and explain them simply. Follow along:
- ▶️ YouTube (daily Shorts + explainers): CD Security Lab
- 💼 LinkedIn: Cuong Do
- 📘 Facebook: CD Security Lab
- 📸 Instagram: @cdsecuritylab
- 📝 More posts like this: Dev.to
Got a question about your own tenant? Drop it in the comments.





Top comments (1)
"Connected means the post office exists" is the right framing, and the Heartbeat check is the right floor. One gap in the prove-it step.
A query over SecurityEvent can only count servers that sent something. A server that has never reported does not show up as a zero. It does not show up at all. So the 15-minute check catches the box that stopped and misses the box that was never wired up, which from your own lab is the failure mode: a DCR with no resources attached means those machines were silent from day one.
Payments has the same shape and a settled answer for it. You cannot find a missing payment by querying your own ledger, because the row is the thing that is missing. You reconcile against the bank's statement. The denominator has to come from somewhere that is not the system under test.
So the alert needs an expected inventory that does not come from telemetry. Asset register, AD computer objects, whatever your source of truth for "this server exists" is, left-joined against Heartbeat, alerting on the empty side. Count the things that should be reporting rather than the things that reported.
The part I would not oversell is the inventory itself. Ours drifts too, and a reconciliation is only ever as good as its denominator. But a drifting asset list fails loudly in a way an empty table never does.