Reading width
Wide uses the full column for everything, text, diagrams, code, and exercises. Narrow keeps the standard reading width.
Text size
Scales the body text. Headings and code blocks keep their size.
In this section
Defender for Endpoint EDR: Alerts, Automation Levels, Isolation, Containment and Indicators
Introduction
The previous lessons were about prevention: controls that refuse something before it happens. Endpoint detection and response is the layer for what prevention misses. It records what devices do, raises alerts on behavior it judges malicious, and can act on a device: remove a file, stop a process, isolate the machine, block an address or contain an account. Whether it does any of that without a person depends on configuration.
This lesson reads what the EDR sensor raised on this estate and what happened after each alert. The sensor saw every stage of the intrusions the earlier lessons met. What followed the alerts is the subject, because on most of them the answer is nothing.
What EDR can do on a device, from least to most intrusive.
The figure is the range of what EDR can do, from keeping a record to cutting a device off the network. Every step past the first is a decision someone made in advance, or a person making it at the time.
Seven Alerts
What the sensor raisedDefender for Endpoint's own alerts carry EDR as their detection source. Joined to the evidence that names the machine each one is about:
AlertInfo
| where DetectionSource == 'EDR'
| join kind=inner (AlertEvidence | where EntityType == 'Machine') on AlertId
| project Day = format_datetime(Timestamp, 'MM-dd HH:mm'), DeviceName, Title
| order by Day asc
The seven alerts, oldest first, with the device each one names:
Day DeviceName Title
----------- -------------- --------------------------------------------------
03-10 10:31 NE-THOMPSON-LT Suspicious process execution
03-10 10:36 NE-THOMPSON-LT Cobalt Strike beacon activity
03-10 10:48 NE-THOMPSON-LT Scheduled task created for persistence
03-12 11:04 NE-BENNETT-LT Ransomware activity detected
03-12 11:06 NE-BENNETT-LT Lateral movement using compromised credentials
03-12 11:12 NE-BENNETT-LT Mass file modification detected
03-12 15:35 WEB-NGE-MCR-01 Anomalous outbound data transfer to known-bad host
Seven alerts on three devices. On 10 March the sensor followed NE-THOMPSON-LT from a suspicious process at 10:31 to beacon traffic at 10:36 and a scheduled task for persistence at 10:48. On 12 March it followed NE-BENNETT-LT from ransomware activity at 11:04 through lateral movement to mass file modification. The same afternoon it raised an outbound transfer from WEB-NGE-MCR-01 to a known-bad host.
Compare these with the antivirus detections of lesson 0.5. The antivirus engine named three files. The EDR sensor described what was happening around them: a beacon, a persistence mechanism, lateral movement, bulk changes to files, data leaving a server. Those are behaviors, and behaviors are what an attacker cannot avoid even when every file they use is new.
That is also why EDR alerts arrive later than antivirus verdicts. A behavior takes time to show: a beacon is a pattern of connections, persistence is a change that has already been made, lateral movement is a logon that has already succeeded.
The sensor has to watch for a while before it can be sure, so an EDR alert is usually a statement about something already under way. That makes what happens next more urgent, not less.
The records beneath the alerts matter as much as the alerts. The sensor keeps process, network, file and registry events whether or not anything alerts, and the investigation of every alert above runs on them. Module 7's telemetry lesson covers what the sensor collects by default and where an estate needs more.
Severity tells the same story. Six of the seven are High, and the scheduled task alert is Medium. A security team reading these at the time would have had no doubt about what they meant.
Notice too what the list does not contain. There are no alerts from the Mac or the Windows servers, and lesson 0.8 asks whether that is because nothing happened there or because the sensor on those devices could not have said so.
The detection layer, in other words, did its job. Everything Module 7 is about starts after that point: what the platform is allowed to do about an alert, and who is expected to do the rest.
Ten Minutes on the Network
The ransomware laptop after its alertLining up NE-BENNETT-LT's records on 12 March shows what the alert was followed by:
union
(DeviceProcessEvents
| where DeviceName == 'NE-BENNETT-LT' and FileName =~ 'notlocker.exe'
| project Timestamp, Record = 'notlocker.exe started'),
(DeviceEvents
| where DeviceName == 'NE-BENNETT-LT' and ActionType == 'AntivirusDetection'
| project Timestamp, Record = 'antivirus detection, not remediated'),
(AlertEvidence
| where DeviceName == 'NE-BENNETT-LT' and EntityType == 'Machine'
| project Timestamp, Record = strcat('EDR alert: ', Title))
| project Time = format_datetime(Timestamp, 'HH:mm:ss'), Record
| order by Time asc
The process start, the antivirus verdict and the three EDR alerts, in time order:
Time Record
-------- ---------------------------------------------------------
11:02:14 notlocker.exe started
11:02:20 antivirus detection, not remediated
11:04:00 EDR alert: Ransomware activity detected
11:06:25 EDR alert: Lateral movement using compromised credentials
11:12:30 EDR alert: Mass file modification detected
The ransomware started at 11:02:14 and the antivirus named it six seconds later without removing it, as lesson 0.5 showed. The EDR alert came at 11:04.
At 11:06 the laptop was using the user's credentials to move to other machines, and at 11:12 it was modifying files in bulk. The ransomware's start and the bulk changes are ten minutes apart, and the laptop stayed on the network for all of it.
Isolation exists for exactly this. Microsoft describes it as disconnecting a compromised device from the network while keeping its connection to the Defender service, so investigation can go on while the device can reach nothing else. An isolated laptop at 11:05 could not have moved laterally at 11:06 or reached shared files at 11:12. Whether that happens on its own, or needs a person, is configuration.
Two settings decide most of it. The automation level of the device group decides whether remediation is taken automatically or waits for approval. Microsoft's page names Full as the recommended level and the default for tenants created since August 2020 with no device groups defined; estates that defined groups keep whatever level they chose.
Automatic attack disruption can isolate end-user workstations on its own when it judges an attack is in progress, a capability Microsoft's page labels a preview. Module 7 covers both, and how to test what an estate is actually set to.
The records cannot show the configuration directly, but they bound it. A laptop that raised a ransomware alert and then moved laterally two minutes later was not isolated in those two minutes, whatever the setting says. If the estate believes it has automatic isolation for workstations, this laptop is the test case that says it did not apply here.
Remediation is the other half. The antivirus verdict on notlocker.exe was not acted on, and nothing in the records shows the EDR layer acting on it either. Under full automation, a file judged malicious is remediated without approval; under the semi levels, a file outside the approved folders waits.
The binary ran from the user's Temp folder, which Microsoft lists as remediated automatically even under the semi level that asks for approval on non-temporary folders, so on this laptop even that level would not have explained the gap.
Four Incidents Nobody Took
What waited for a personThe alerts above were grouped into incidents. Their state, taken from the latest record for each:
SecurityIncident
| where IncidentNumber in (100448, 100468, 100472, 100485)
| summarize arg_max(TimeGenerated, *), Updates = count() by IncidentNumber
| project IncidentNumber, Severity, Status, Updates, Owner
The four incidents, each with its latest status:
IncidentNumber Severity Status Updates Owner
-------------- -------- ------ ------- -----
100485 High New 1 {}
100472 High New 1 {}
100468 High New 1 {}
100448 High New 1 {}
Four High-severity incidents, all New, each with exactly one record: created and never touched. No owner was assigned and the status never changed. The earliest is the Thompson intrusion on 10 March; the latest is the web server on 12 March. Lesson 0.1 met the first as the intrusion it walked through.
That is the context the configuration has to be judged in. An estate whose incidents are worked within minutes can afford to keep the automation level low, because a person will isolate the laptop. An estate whose incidents sit for days cannot, because every action that waits for a person waits for days. The records show which estate this is.
Read the incident list the way a manager would. Four High-severity incidents in three days, on four hosts, each a separate stage of what may be a single campaign. Each one, opened, would have led an analyst to the next.
A queue that leaves those untouched is either not being watched or is being watched by people with no time to open them. Either way, the alerts are reaching nobody who can act, and the platform is the only responder available.
None of this is a criticism of the analysts. It is a fact about capacity, and capacity is the input that decides how much the platform should do without asking. A team of two covering business hours needs a platform that handles the night; a team running shifts can keep more decisions for itself. Neither is wrong, but the configuration has to match the team that actually exists.
Module 7's lesson on choosing the automation level starts from exactly this question: who is going to act, and how quickly, on the alerts that arrive at three in the morning.
The incidents also show the cost of an alert nobody owns. The Thompson incident was the earliest warning of the month, raised two days before the ransomware. Whatever came later on this estate came after a High-severity incident had been waiting, unread, for two days.
Module 7 does not fix the queue; that is an operations question. What it does is make the platform's own response match the queue the estate actually has, so that the actions which cannot wait do not have to.
The Response That Did Act
Automatic containment of an accountOne response on this estate did happen without a person. Overnight on 12 March, a lateral movement chain used the svc-fileshare account to reach three hosts. The records of what followed come from a different table:
DisruptionAndResponseEvents
| project Time = format_datetime(Timestamp, 'HH:mm:ss'), DeviceName, ActionType,
Account = SourceUserName, From = SourceDeviceName
| order by Time asc
The seven records of refused access, in time order:
Time DeviceName ActionType Account From
-------- ------------ ------------------------------- ------------- -----------
02:34:11 NE-KHAN-LT ContainedUserSmbFileOpenBlocked svc-fileshare NE-LEWIS-LT
02:34:46 NE-KHAN-LT ContainedUserLogonBlocked svc-fileshare NE-LEWIS-LT
02:36:20 NE-KHAN-LT ContainedUserRpcAccessBlocked svc-fileshare NE-LEWIS-LT
02:37:05 NE-PATEL-LT ContainedUserSmbFileOpenBlocked svc-fileshare NE-LEWIS-LT
02:37:41 NE-PATEL-LT ContainedUserLogonBlocked svc-fileshare NE-LEWIS-LT
02:41:02 NE-SHARMA-LT ContainedUserSmbFileOpenBlocked svc-fileshare NE-LEWIS-LT
02:41:30 NE-SHARMA-LT ContainedUserLogonBlocked svc-fileshare NE-LEWIS-LT
Between 02:34 and 02:41, the svc-fileshare account, coming from NE-LEWIS-LT, tried to reach NE-KHAN-LT, NE-PATEL-LT, NE-SHARMA-LT over SMB, network logon and RPC, and each laptop refused it.
These records are the outcome of the contain-user action that automatic attack disruption took on the account after its alert. Microsoft documents contain user as one of the containment actions alongside isolation, and the table records what onboarded devices refused once it was in place.
The contrast with the rest of this lesson is the point. The platform judged one account compromised and acted on it within minutes, at night, with nobody watching, and three laptops were spared.
On the same estate, the ransomware laptop was not isolated, and the web server's outbound address was not blocked. The difference is not the quality of the detection; every one of those was alerted. It is which response the estate allowed the platform to take on its own.
The timing is the other lesson in these records. The account's alert, raised by Defender for Identity rather than the endpoint sensor, came at 02:32, after its hops between hosts, and the first refusal followed at 02:34. Containment did not stop the hops that had already happened; it stopped the ones that came next.
Every automatic response works this way. It cannot undo what happened before the alert, only shorten what happens after it, which is why the minutes between an alert and a response matter more than the quality of the alert itself.
Containment has limits worth knowing, too. It acts through onboarded devices, which refuse the contained account or address. A device with no sensor cannot refuse anything.
NE-FOSTER-WS, the chain's second hop, has no inventory record at all, so it could not have refused the account even with containment in place. Coverage and response depend on each other. A contained account is refused by every device that has a sensor to refuse it, and by nothing else.
Two Days After the Alert
An address that was never blockedThe web server's alert on 12 March named 45.137.21.88 as the known-bad host. Every connection the server made to that address, split at the moment of the alert:
DeviceNetworkEvents
| where DeviceName == 'WEB-NGE-MCR-01' and RemoteIP == '45.137.21.88'
| summarize
BeforeAlert = countif(Timestamp < datetime(2026-03-12 15:35)),
AfterAlert = countif(Timestamp >= datetime(2026-03-12 15:35)),
Last = format_datetime(max(Timestamp), 'MM-dd HH:mm')
The connections before and after the alert, with the last one:
BeforeAlert AfterAlert Last
----------- ---------- -----------
19 16 03-14 15:45
19 connections before the alert and 16 after it, the last at 15:45 on 14 March, seventeen minutes before the coin miner from lesson 0.5 was detected. The address was known, named in a High-severity alert and owned by nobody. The server kept reaching it for two days.
An IP indicator is the narrowest response EDR offers: a block on one address, on every onboarded device, without touching anything else. Microsoft's indicator page lists Linux among the supported platforms and says a block can take up to 48 hours to reach a device, with most taking effect in under two.
An indicator created when the alert fired at 15:35 on 12 March would, even at the documented worst case, have been in force before the last connection at 15:45 on 14 March; in the usual case it would have stopped all but the first two hours of them.
Indicators are cheap, reversible and precise, which makes them the response to configure first. They are also the one this estate most visibly did not use.
Module 7's indicator lesson covers creating them from alerts, expiring them, and the conditions that make one silently do nothing, such as network protection left in audit on the devices meant to enforce it. Lesson 0.6 showed that exact condition on this estate before 11 March.
The web server makes a sharper point about choosing a response. Isolating a production web server stops the business, and an analyst might reasonably hesitate. Blocking one outbound address it has no business reaching stops nothing the business needs. The cheapest response was the most appropriate one, and it was the one available for longest without being used.
Indicators also outlive the incident. The same address blocked on every onboarded device protects the laptops too, and any other server the same attacker reaches later. A response taken on one alert becomes prevention for the whole estate, which is a better return than any single isolation can give.
What Module 7 Builds On It
EDR configuration and automated responseThe estate's EDR position fits on one card:
Three hosts, one pattern
Alerted is not respondedModule 7 begins with what EDR does beyond its defaults and with automatic remediation, then the choice of automation level per device group. It covers live response, both configuring it and scripting it, isolation and containment, and indicator management. It ends with alert tuning, threat analytics, the telemetry the sensor collects, and one response seen through from alert to closed incident.
The order follows this lesson's findings. Automation level comes early because it decides what happens to every alert before anyone reads it. Live response comes next because it is how an engineer acts on a device the platform did not act on.
Isolation, containment and indicators are the three responses this lesson found missing or working. Tuning and telemetry come last, because they make the alerts that drive everything else fewer and better, and that only matters once something is acting on them.
What Module 7 builds
Module 7The habits to carry into it are the ones this lesson used. Read what followed an alert, not only the alert. Line a device's records up after its alert to see how long it stayed able to act. Check incidents for an owner and a change of status, because an untouched incident is a response that never started.
Find the responses that did happen without a person, because they show what the estate already allows. And split an address's connections at its alert, because every one after the alert is a block that was available and unused.
None of these habits needs the console. They need the alert tables, the incident table, the disruption table and the device tables, all of which an engineer can query, and all of which tell the truth about what the configuration did whatever the configuration pages say.
Each of those is a question about configuration answered from records rather than from a settings page, and the answers here all point the same way. This estate detects well and responds when the platform is permitted to. Module 7 is about widening that permission deliberately, one response at a time, with the records as the test.
Practice
Decide where the response belonged prepared evidenceThe prepared evidence below is the seven EDR alerts with the devices they name. Decide which alert most justified isolating a device at once, which named an address that should have become an indicator, and what the night's automatic containment says about the rest. Each answer comes with its reasoning, so check it against the rows rather than only the choice.
Read the alerts in order, as a responder would have met them. The question at each one is not whether it was right, since all of them were, but which response was available at that moment and whether anything was configured to take it.
Keep the three responses distinct as you read. Remediation acts on a file or process on one device. Isolation cuts one device off from everything except the Defender service. An indicator blocks one address or file everywhere at once. Each alert above has one of those as its natural response, and choosing the right one is most of the skill.
The incident list beside them matters as well: an alert whose incident was never opened is one where no response was chosen at all, by the platform or by a person.
Then run the web server query on this page with the split moved to two hours after the alert, the usual time for an indicator to take effect, and count how many connections an indicator would still have allowed. Finally, rerun the incident query for every incident on the estate rather than these four, and count how many of the High-severity ones were ever assigned to anyone.
EDR saw every stage of this estate's intrusions and alerted on each of them. Where the platform was allowed to act on its own it did, within minutes. Everywhere else the alert waited for a person, and the records show nobody came: a laptop left on the network for ten minutes of ransomware and a server left reaching a known-bad address for two days.
The next lesson turns to the devices that are not Windows laptops, the servers, the Linux web server and the Mac, and what each needs from every control in this module.