Lesson 61 — Introduction to Advanced Threat Hunting in Microsoft Sentinel
Move beyond reacting to alerts. Learn to form hypotheses, search available telemetry, test alternative explanations and document defensible findings in Microsoft Sentinel.

What you will learn
Think like a threat hunter, not an alert counter.
Agent Foskett case file — The alert that never fired
There is no high-severity Sentinel incident. Yet an analyst notices a service account authenticating at unusual hours. Is this an expected batch job, a maintenance change or early evidence of misuse? Agent Foskett begins with a question, not a verdict.
What advanced threat hunting means
Threat hunting is a proactive, hypothesis-driven search for suspicious activity that may not have generated an alert. It differs from alert triage: hunters choose a question, inspect relevant evidence, compare behavior against a baseline and decide whether further investigation is justified.
Learning objectives
- Distinguish proactive hunting from incident response and scheduled detection.
- Translate a suspicion into a testable hypothesis.
- Check which telemetry is actually ingested and retained.
- Use KQL to scope, summarize and investigate candidate anomalies.
- Document negative findings, coverage limits and follow-up pivots.
The evidence-first hunting cycle
Prerequisites and lab safety
- Complete the earlier Sentinel KQL and investigation lessons or have equivalent experience.
- Use an authorized Microsoft Sentinel workspace and a controlled lab account.
- Confirm Entra
SigninLogsis ingested before running the examples. - Have permission to read logs and document a hunt.
- Keep this lesson read-only: do not disable accounts or block addresses based on a single anomaly.
Step 1 — Start with a hypothesis
Example: “A service account may be signing in interactively outside its normal operating window.” This is testable. The statement “The account is compromised” is a conclusion and must not be used as the starting assumption.
Step 2 — List competing explanations
- Approved overnight maintenance or scheduled processing.
- A legitimate application deployment or service migration.
- Clock, time-zone or ingestion effects.
- Unexpected use of credentials by an operator.
- Unauthorized access requiring deeper investigation.
Write these down before viewing the results to reduce confirmation bias.
Step 3 — Identify relevant data sources
| Question | Potential source | Limitation |
|---|---|---|
| When did an account sign in? | SigninLogs | Requires Entra diagnostic ingestion |
| What identity changes occurred? | AuditLogs | Audit operation context varies |
| What did a device execute? | Defender XDR advanced hunting | Different query surface and permissions |
| What alerts were produced? | Sentinel incidents/alerts | Absence of alert is not absence of activity |
Step 4 — Check coverage before hunting
Verify the relevant table exists, has recent data and covers the intended users and time range. A query returning no records could reflect a data collection gap rather than an absence of suspicious behavior.
Step 5 — KQL: inspect sign-in coverage
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
SigninLogs
| where TimeGenerated >= ago(7d)
| summarize Events = count(),
FirstSeen = min(TimeGenerated),
LastSeen = max(TimeGenerated),
Accounts = dcount(UserPrincipalName)
by bin(TimeGenerated, 1d)
| order by TimeGenerated ascRun this in the Sentinel Log Analytics workspace where SigninLogs is available. Look for missing days or sudden drops before drawing conclusions.
Step 6 — Establish a sensible baseline
Compare recent behavior with a longer, comparable period. Account for working days, maintenance windows, shift patterns and time zones. A deviation is a lead; it is not proof of malicious activity.
Step 7 — KQL: find accounts with unusual sign-in hours
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
- 12
- 13
- 14
let Lookback = 7d;
SigninLogs
| where TimeGenerated >= ago(Lookback)
| where isnotempty(UserPrincipalName)
| extend HourUTC = datetime_part("hour", TimeGenerated)
| where HourUTC between (0 .. 5)
| summarize Attempts = count(),
DistinctIPs = dcount(IPAddress),
FirstSeen = min(TimeGenerated),
LastSeen = max(TimeGenerated)
by UserPrincipalName
| where Attempts >= 3
| order by DistinctIPs desc, Attempts descThe hours are UTC, not necessarily local business hours. The threshold is an illustrative filter; validate expected schedules and account types before interpreting results.
Step 8 — Why a hunting result is not an incident
An account appearing in the results simply met the query conditions. Check approved activity, device context, source IP ownership and sign-in outcome. Record whether the hypothesis was supported, contradicted or remained inconclusive.
Step 9 — Pivot from account to IP and application
Choose one authorized lab account and examine its IP addresses, applications, sign-in outcomes and timing. Avoid immediately expanding to the whole tenant when a narrow pivot can answer the question more safely.
Step 10 — KQL: build an account investigation timeline
- 1
- 2
- 3
- 4
- 5
- 6
- 7
- 8
- 9
- 10
- 11
let TargetUPN = "analyst@example.com";
SigninLogs
| where TimeGenerated >= ago(7d)
| where UserPrincipalName =~ TargetUPN
| project TimeGenerated, UserPrincipalName,
IPAddress, AppDisplayName,
ResultType, ConditionalAccessStatus,
CorrelationId
| order by TimeGenerated asc
| take 200Replace the example UPN with an approved lab account. The row cap may truncate busy accounts; note that limitation in the hunt record.
Step 11 — Enrich candidates with approved context
Review asset ownership, account purpose, documented maintenance, known VPN egress and approved identity context. Do not treat a trusted IP label as a verdict. If using threat intelligence, record its source and observation date.
Step 12 — Separate event time from ingestion time
Use TimeGenerated consistently for event-based analysis, but remember ingestion delays can affect near-real-time completeness. Document when the hunt ran and the time window examined.
Step 13 — Hunting versus detections
A hunt is an exploratory investigation. If repeated testing produces a reliable signal, consider converting it into an analytic rule after validating false positives, scheduling, entity mapping and alert fatigue. Do not promote a one-off anomaly query directly into production detection.
Step 14 — Keep a hunt notebook
Step 15 — Avoid common analytical traps
- Confirmation bias: seeking only evidence that supports the first suspicion.
- Survivorship bias: inspecting only devices or accounts with full telemetry.
- Base-rate neglect: treating rare but legitimate behavior as inherently hostile.
- Time-zone errors: interpreting UTC as local operating time.
- False precision: claiming “no compromise” from a bounded query.
Step 16 — Define escalation criteria
Escalate when corroborating evidence indicates material risk, such as suspicious authentication coupled with unexplained privilege changes or endpoint activity. Follow the organization’s incident process and preserve query results, identifiers and timestamps.
Step 17 — Work within permissions and privacy
Use least-privilege log access, restrict exposure of user identifiers and document why each data source is needed. Hunting is not permission to export unrestricted identity or endpoint telemetry.
Step 18 — Cost and query performance
Start with a bounded time window, filter early, summarize before joining large tables and project only needed columns. Wide, unbounded queries can increase cost and produce misleadingly incomplete results due to limits.
Step 19 — Build a repeatable hunt
Save the hypothesis, parameters, query version, source dependencies and review criteria. A useful hunt can be repeated with updated data without quietly changing the investigative question.
Agent Foskett case file — The service account after midnight
The initial query highlights several overnight sign-ins. Investigation shows most belong to approved scheduled jobs, but one account is being used with an unexpected application. Agent Foskett records the anomalous application and asks the service owner to verify the change before escalating.
Finding: the hunt did not prove compromise. It narrowed the uncertainty to a specific, testable follow-up.
Hands-on lab — your first hypothesis-driven hunt
- Write one suspicious sign-in hypothesis and two alternative explanations.
- Verify seven days of
SigninLogscoverage. - Run the off-hours candidate query.
- Select a permitted test account and build its timeline.
- Check whether the account behavior matches approved operations.
- Record evidence, gaps and a follow-up pivot.
- Decide whether to close the hunt, repeat it or escalate.
Knowledge check
- Why is a testable hypothesis better than beginning with “compromise confirmed”?
- What does a zero-row query actually prove?
- Why must UTC hours be interpreted carefully?
- When is a hunt ready to become an analytic rule?
- What evidence should accompany an escalation?
Publishing and quality checklist
- Lesson title, Module 7 label and canonical URL correct.
- Opening GEMXIT HTML comment retained before doctype.
- Original background, footer, scripts and navigation preserved.
- All three KQL editors full width with visible line numbers.
- Queries use actual Sentinel Logs tables and bounded windows.
- Missing telemetry and alternative explanations acknowledged.
- Previous lesson linked; unpublished next lesson not linked.
Agent Foskett takeaway
A threat hunter does not start by proving an attacker exists. A threat hunter starts by asking a question the logs can answer. Build the hypothesis, test it fairly, document what you can and cannot know, and follow the next piece of evidence.
The Logs Already Knew.
