Threat hunting is the practice of searching an organization's systems for signs of an attacker that automated alerts have not caught. Instead of waiting for an alarm, an analyst starts from a question, such as "could an attacker be doing this here?", and looks through the logs for evidence. TaHiTI, an open hunting method written by Dutch financial institutions, defines it as a proactive search for malicious activity, current and historical, that has evaded existing security defenses. This article explains how hunting works and how a small team can start.

Hunting versus alert-driven monitoring

Tools such as a SIEM (a system that collects and analyses logs from many sources) and EDR (software that watches laptops and servers and responds to threats) raise an alert when a rule or model matches. The team then reacts to that alert. Hunting works in the opposite direction. A person chooses a question first, searches the data, and then decides whether something is wrong. TaHiTI says hunting assumes a security monitoring process is already running properly, so hunting adds to monitoring and does not replace it. For how these tools relate, see SIEM, SOAR and XDR compared and EDR, XDR and MDR compared.

TaHiTI gives the main purpose of hunting as shortening the time an attacker stays unnoticed after getting in. This time is called dwell time. It lists other benefits as well:

  • Finding gaps in visibility, meaning data you would need to see a specific attacker technique but do not collect
  • Finding gaps in detection, meaning behavior that no existing rule would catch
  • Creating ideas for new monitoring use cases and detection rules
  • Uncovering new threats and attacker techniques that feed back into threat intelligence
  • Producing recommendations for new preventive measures

Three ways to run a hunt

PEAK is a hunting framework introduced by David Bianco on Splunk's blog in April 2023. Splunk created it, so it is a vendor's framework, though it is not tied to one product. It describes three types of hunt:

  • Hypothesis-driven: form a theory about a likely threat, then use data to confirm or reject it
  • Baseline: define what normal behavior looks like, then investigate deviations
  • Model-assisted: use machine learning models of known-good or known-malicious behavior to find activity that matches or deviates

TaHiTI divides hunting more simply, into unstructured hunting (digging through data for anomalies, with no hypothesis) and structured hunting (starting from a hypothesis). In our view, a beginner team should start with hypothesis-driven hunts. They have a clear scope, they end with an answer, and the result is easy to write down.

Write a hypothesis from attacker behavior

A hypothesis is a statement that data can prove or disprove. MITRE ATT&CK, a public knowledge base of attacker behavior built from real-world observations, is a good source of ideas. It describes tactics (what the attacker is trying to achieve) and techniques (how the attacker does it). TaHiTI lists ATT&CK as one of the triggers that can start a hunt.

Here is our own example. Hypothesis: an attacker who took over an ordinary user account is using it to log in to many servers that the account has never used before. Data needed: login logs from the servers. Check: list the accounts that reached an unusual number of servers in a short time, then ask each account owner whether it was them. TaHiTI describes three possible outcomes: proven (malicious activity found, so incident response starts), disproven (nothing found), or inconclusive (change the scope or data and try again).

Prefer hunting for behavior over hunting for single indicators. The Pyramid of Pain, a model by David Bianco that TaHiTI also uses, shows why. Its six layers from bottom to top are hash values (a fingerprint of a file), IP addresses, domain names, network and host artifacts (traces left on the network or machine), tools, and TTPs (tactics, techniques and procedures, the attacker's behavior). Bianco rates how hard each layer is for an attacker to change: trivial, low, low to moderate, moderate, high, and very high. TaHiTI says it focuses on the top three layers, with the most focus on TTPs.

INFOGRAPHIC
Hunting indicators versus behavior

Easy for attacker to change (bottom 3)

  • File hashtrivial to change
  • IP addresslow difficulty
  • Domain namelow to moderate difficulty

Hard for attacker to change (top 3)

  • Network and host artifactsmoderate difficulty
  • Toolshigh difficulty
  • TTPs (behavior)very high difficulty

How a hunt runs: TaHiTI in three phases

TaHiTI describes a hunt in three phases and six steps in all. PEAK has a similar shape with its own three phases: Prepare, Execute and Act. Act covers documentation, automation and communication. TaHiTI's three phases are:

  • Initiate: a trigger, such as threat intelligence or ATT&CK, starts the process. The team turns it into a short abstract and stores it on a hunting backlog (a queue of hunts waiting to be done). A simple tool such as SharePoint or JIRA is enough
  • Hunt: the team defines the hunt, writes the hypothesis, and picks data sources and analysis techniques. It analyses the data, refines the scope as needed, then validates the hypothesis as proven, disproven or inconclusive. If the data needed does not exist, the hunt counts as failed, but TaHiTI notes it can still give useful lessons
  • Finalize: the team documents the results and conclusions, with recommendations such as preventive measures, logging changes, new monitoring use cases and process improvements, then hands the results to other processes
INFOGRAPHIC
TaHiTI in three phases
01
Initiate
trigger, then backlog
02
Hunt
test hypothesis on data
03
Finalize
document, hand over

The last step is a handover. TaHiTI says the results can go to security incident response, security monitoring, threat intelligence, vulnerability management and other processes. These same processes also supply triggers for new hunts, so hunting can speed up improvement in all of them.

INFOGRAPHIC
Where one hunt's results can go
Results of one hunt
handed to other processes
Incident response
when malicious activity is found
Security monitoring
new use cases and rules
Threat intelligence
newly found attacker behavior
Vulnerability management
weaknesses and misconfigurations

What you need before you start

A SANS paper by Robert M. Lee and Rob T. Lee (2016) lists what effective hunting needs:

  • Enough data to search, such as logs, flow records (records of network connections) and alerts
  • Tools for searching, visualization and analysis
  • Analysts with curiosity and judgment, because repetitive tasks can be automated but the thinking cannot
  • Dedicated time, because hunters work best when they are not also handling alert queues or maintenance

The same paper describes the Hunting Maturity Model, created by David Bianco, which has five levels. Level 0 relies on automated alerts. Level 4 automates most of the procedures that worked. Hunting starts at level 1 with searches for threat intelligence indicators. In our view, a small team does not need to aim high at first, and one repeatable hunt a month is a reasonable start. Without logs there is nothing to hunt, so read What to Log and How Long to Keep It first. Threat intelligence is also a rich source of hypotheses; see Cyber Threat Intelligence for Beginners.

How to tell whether hunting is worth it

TaHiTI suggests metrics that show the value hunting adds, rather than only counting hunts. Examples are:

  • The dwell time of any compromise the hunt uncovered
  • The number of incidents that hunting triggered
  • The number of monitoring use cases added or updated
  • New threat intelligence created during hunting
  • New preventive measures recommended in hunting reports
  • The number of vulnerabilities or misconfigurations uncovered

Common mistakes

  • Hunting with no hypothesis: TaHiTI warns that a badly defined hypothesis leads to no results, or to wrong results and wrong conclusions
  • Treating hunting as a replacement for monitoring: TaHiTI assumes monitoring already works, and hunting is meant to find what it missed
  • Hunting where the data does not exist: in our view, logs that were never collected or were deleted too early make the hunt fail before it starts
  • Searching only for known hashes and IP addresses: in the Pyramid of Pain these are the layers an attacker changes most easily
  • Not writing down a hunt that found nothing: a disproven hypothesis, the data used and the gaps found are still useful results
  • Using ATT&CK as a checklist to reach 100% coverage: MITRE's own page cautions against this
  • Leaving findings in the report: no new rule, log source or fix follows, so the next attacker is missed the same way

Start today

  • Pick one hypothesis: use the example in this article, or one ATT&CK technique that fits your systems
  • Check the data first: confirm that the logs you need exist and cover enough days; if they do not, fixing the logging is your first result
  • Start a backlog: a spreadsheet or SharePoint list is enough, with each idea written as a short abstract
  • Write every result in a fixed format: hypothesis, data used, method, outcome (proven, disproven or inconclusive) and recommendations
  • Hand every finding to an owner: open a ticket for the monitoring, incident response or vulnerability team

To see where your detection and response capability stands today, try the self-assessment.

Scope and accuracy

This article draws on the pages listed below, which we opened in November 2026. We read TaHiTI's short public whitepaper, not its full documentation or its MaGMa tool. The SANS Hunting Maturity Model page did not describe the levels, so the level descriptions come from the 2016 SANS paper. The PEAK description comes from Splunk's blog post. ATT&CK v20 was the version announced on MITRE's site when we read it, so check for newer versions. The hypothesis example and our views on where to start are our own and are not from these sources.

References and further reading