Advanced Hunting in Defender XDR | Guided by A Guide to Cloud

What is Advanced Hunting?

Simple explanation

Advanced Hunting is your search engine for security data in Defender XDR. You write KQL queries and get answers across endpoints, emails, identities, and cloud apps — all from one place.

Unlike analytics rules (which detect known patterns automatically), hunting is proactive. You are the one asking the questions: “Has anyone in the org been contacted by this phishing domain?” or “Are there any devices running processes from temp folders?”

When your hunting query finds something valuable, you can save it as a custom detection rule (Module 9) so it runs automatically from now on.

Hunting patterns: real-world queries

Pattern 1: Living off the land (LOLBins)

Attackers use legitimate Windows tools for malicious purposes. Hunt for unusual parent-child process relationships:

DeviceProcessEvents
| where Timestamp > ago(7d)
| where InitiatingProcessFileName in ("winword.exe", "excel.exe", "powerpnt.exe")
| where FileName in ("powershell.exe", "cmd.exe", "mshta.exe", "wscript.exe", "cscript.exe")
| project Timestamp, DeviceName, AccountName, InitiatingProcessFileName, FileName, ProcessCommandLine
| sort by Timestamp desc

What this finds: Office applications spawning command interpreters — a strong indicator of macro-based malware.

Pattern 2: Credential theft indicators

DeviceProcessEvents
| where Timestamp > ago(24h)
| where FileName in ("mimikatz.exe", "procdump.exe", "sekurlsa.exe")
    or ProcessCommandLine has_any ("sekurlsa", "lsadump", "kerberos::list", "privilege::debug")
| project Timestamp, DeviceName, AccountName, FileName, ProcessCommandLine

What this finds: Known credential dumping tools and command-line indicators.

Pattern 3: Phishing campaign scope

EmailEvents
| where Timestamp > ago(48h)
| where SenderFromAddress =~ "hr-updates@pacificmeridian-careers.com"
| summarize
    TotalEmails = count(),
    Recipients = make_set(RecipientEmailAddress),
    Delivered = countif(DeliveryAction == "Delivered"),
    Blocked = countif(DeliveryAction == "Blocked")
| extend RecipientCount = array_length(Recipients)

What this finds: The full scope of a phishing campaign — how many emails, to whom, and how many got through.

Pattern 4: Lateral movement via RDP

DeviceLogonEvents
| where Timestamp > ago(7d)
| where LogonType == "RemoteInteractive"
| where IsLocalAdmin == true
| summarize
    RDPSessions = count(),
    SourceDevices = make_set(RemoteDeviceName)
    by DeviceName, AccountName
| where RDPSessions > 5
| sort by RDPSessions desc

What this finds: Devices with unusually high admin RDP sessions — a lateral movement indicator.

Pattern 5: Cross-domain hunting (email to endpoint)

let PhishingUrls =
    EmailUrlInfo
    | where Timestamp > ago(48h)
    | where UrlDomain == "evil-login.com"
    | distinct Url;
DeviceNetworkEvents
| where Timestamp > ago(48h)
| where RemoteUrl in (PhishingUrls)
| project Timestamp, DeviceName, AccountName, RemoteUrl

What this finds: Devices that visited a phishing URL found in emails — connecting the email attack vector to endpoint compromise.

Exam tip: cross-table joins

Cross-domain hunting (combining email + endpoint, or identity + endpoint data) is a key exam topic. The let statement creates a variable holding results from one table, which you then use to filter another table.

This is the power of Defender XDR’s unified hunting — you can trace an attack from the phishing email through the endpoint compromise in a single query.

The hunting workflow

Step	What You Do	Example
1. Hypothesis	Form a question based on threat intel, incident learning, or MITRE technique	”Are there signs of Kerberoasting in our environment?“
2. Query	Write KQL to search for evidence	Query IdentityQueryEvents for abnormal TGS requests
3. Analyse	Review results — separate signal from noise	Filter out service accounts that normally make TGS requests
4. Refine	Narrow the query to reduce false positives	Add exclusions for known legitimate behaviour
5. Detect	Save as custom detection rule if the hunt finds a repeatable pattern	Create hourly detection for Kerberoasting indicators

Scenario: Tyler hunts for data exfiltration at CipherStack

Tyler hypothesises that an attacker may be exfiltrating code via DNS tunnelling (encoding data in DNS queries to bypass firewalls).

Hunting query:

DeviceNetworkEvents
| where Timestamp > ago(7d)
| where RemotePort == 53
| extend DomainLength = strlen(RemoteUrl)
| where DomainLength > 50  // DNS tunnelling uses very long domain names
| summarize
    LongDnsQueries = count(),
    AvgLength = avg(DomainLength),
    SampleDomains = make_set(RemoteUrl, 5)
    by DeviceName
| where LongDnsQueries > 100
| sort by LongDnsQueries desc

Results: One developer workstation made 2,300 DNS queries with domain names averaging 85 characters — classic DNS tunnelling. Tyler saves the query as a custom detection rule running every 3 hours.

Question

What is the hunting workflow in 5 steps?

Click or press Enter to reveal answer

Answer

1. Hypothesis — form a question (threat intel, MITRE, incident learning). 2. Query — write KQL to search. 3. Analyse — review results, separate signal from noise. 4. Refine — narrow query, reduce false positives. 5. Detect — save as custom detection rule for ongoing monitoring.

Click to flip back

Question

How do you trace a phishing attack from email to endpoint in one query?

Click or press Enter to reveal answer

Answer

Use a 'let' statement to extract phishing URLs from EmailUrlInfo, then search DeviceNetworkEvents for devices that visited those URLs. This cross-table approach connects the email attack vector to the endpoint compromise.

Click to flip back

Question

What KQL pattern detects Office apps spawning command interpreters (macro malware)?

Click or press Enter to reveal answer

Answer

DeviceProcessEvents where InitiatingProcessFileName in ('winword.exe', 'excel.exe') and FileName in ('powershell.exe', 'cmd.exe', 'mshta.exe'). This parent-child relationship is a strong indicator of macro-based malware delivery.

Click to flip back

Knowledge Check

Tyler wants to find all devices in CipherStack's environment that made DNS queries with unusually long domain names (potential DNS tunnelling). Which table and filter should he use?

Knowledge Check

Anika wants to trace which devices visited phishing URLs that arrived via email. She needs to combine EmailUrlInfo and DeviceNetworkEvents. What KQL technique should she use?

Knowledge Check

Tyler discovers suspicious DNS activity on a developer workstation. His initial KQL query shows 2,300 DNS queries with domain names averaging 85 characters. He suspects DNS tunnelling but needs to confirm. The legitimate CDN the company uses also generates long subdomains. What is the BEST next step?

🎬 Video coming soon

Next up: Advanced Hunting queries are written. Now let’s build Sentinel-specific hunting queries and learn how to monitor hunts over time.