Skip to main content
A hand lens resting on a stack of printed log paper, standing for a hunter searching telemetry for missed attacker activity

What is threat hunting?

Threat hunting is a proactive, hypothesis-led search of security telemetry for attackers your detections missed. Hunt types, maturity models, and what a hunt produces.

Explainer
Updated 9 min read

Definition

Threat hunting

Also called Cyber threat hunting, Proactive threat hunting, Hunt operations

Threat hunting is the proactive search through an organization's security telemetry for attacker activity that its automated detections missed, starting from a hypothesis, a piece of threat intelligence, or a known attacker technique rather than from an alert. A hunt ends with an answer to its question, and its lasting outputs are new or better detections, closed visibility gaps, and incidents opened for anything it found.

Hunting teams with a formally defined methodology (SANS, 2026)
37%
Teams that formally measure hunt outcomes (SANS, 2026)
40%
Most common length of a single hunt, reported by 31% (SANS, 2026)
9-24 hours
Name data quality or quantity as their top barrier (SANS, 2026)
50%

How a threat hunt works

A threat hunt starts from a question, not an alert. Definitions vary in wording but agree on the core. Sqrrl's hunting framework calls it "the process of proactively and iteratively searching through networks to detect and isolate advanced threats that evade existing security solutions." MITRE's TTP-Based Hunting report defines it as "the proactive detection and investigation of malicious activity within a network." David Bianco, who leads Splunk's PEAK framework, puts the purpose most plainly: "any manual or machine-assisted process intended to find security incidents missed by an organization's automated detection systems."

Hunts differ mainly in where the question comes from:

  • Hypothesis-driven. The hunter states a testable guess about attacker behavior and tries to prove or disprove it. PEAK's worked example narrows "data exfiltration" to "A threat actor may be exfiltrating sensitive financial data using DNS tunneling," and uses its ABLE checklist to pin down the Actor, Behavior, Location, and Evidence (which data sources would show it, and what it would look like).
  • Intel-driven. A threat report, campaign write-up, or new vulnerability supplies the lead. The hunter extracts behaviors and indicators and searches historical data for them. Bianco's maturity model describes this as the first kind of hunting most teams do.
  • TTP-driven, or threat-model-driven. The hunter works from the techniques an attacker would need in this environment, usually mapped to MITRE ATT&CK. MITRE's report argues this holds up better than indicator sweeps because the operating system constrains which techniques an attacker can use after compromise, and "there are a relatively small number of these techniques."

PEAK, which Splunk's SURGe team introduced in April 2023, adds two more types: baseline hunts (also called exploratory data analysis), which profile normal activity in a data source and look for deviations, and model-assisted threat hunts, which use machine learning. All PEAK hunts run in three phases: Prepare (pick the topic and plan), Execute (gather and analyze data), and Act (document, automate, and communicate).

How hunting differs from detection and alert triage

Hunting, detection, and alert triage share data and often people, but they start from different places and produce different things.

Threat huntingDetectionAlert triage
Starts fromA hypothesis, intel, or techniqueA rule written in advanceAn alert that already fired
RunsTime-boxed, on demand or scheduledContinuously, on a scheduleEach time an alert arrives
Question askedIs this happening anywhere?Did this pattern just match?Is this one alert real?
Main outputFindings, new detections, gapsAlertsA verdict: close or escalate

Detection only finds what someone already knew to write a rule for. Triage only sees what detection raised, which is why a noisy queue causes alert fatigue without making coverage any better. Hunting is the loop that finds what neither covers, then hands the method to detection so the next occurrence raises an alert on its own. Sqrrl described that hand-off as the last stage of its hunting loop: "successful hunts form the basis for informing and enriching automated analytics."

Threat hunting maturity models

Two models are in common use, and they answer different questions.

The Hunting Maturity Model (HMM) rates a whole program. Bianco first published it in 2015. It runs from HMM0, where a team relies on automated alerting and "is not considered to be capable of hunting," to HMM4, where "any successful hunting process will be operationalized and turned into automated detection." In between, HMM1 searches historical data for indicators from threat reports, HMM2 runs hunt procedures written by others, and HMM3 creates its own. Bianco called HMM2 the most common level among teams with active hunting programs. The model's warning about HMM3 is the one most teams hit: as procedures pile up, running them all on schedule takes more analysts, which is the reason HMM4 automates successful hunts. Agent-run hunting is a newer way past that limit: an agent can rerun a procedure on schedule and draft the resulting rule, as in the GitHub example under Examples, while a person still reviews each rule before it ships.

PEAK structures individual hunts rather than grading programs. It also borrows the HMM for program maturity and recommends that most hunting organizations reach at least HMM2.

TaHiTI (Targeted Hunting integrating Threat Intelligence), created by the Dutch Payments Association in 2018, is a third methodology. The SANS 2026 Threat Hunting Survey recommends anchoring hunts in "a repeatable framework like PEAK or TaHiTI."

The survey suggests most programs aren't there yet. Of 500 respondents, 37% have a formally defined hunting methodology, down from 46% in 2025 and 51% in 2024, while ad hoc approaches climbed to 39%. Only 40% formally measure hunt outcomes, down from 64% in 2024. A typical hunt is not quick: 31% said one takes nine to 24 hours, and 24% finish within a full day. For the first time in the survey's history, data quality or quantity (50%) outranked skilled staff (45%) as the top barrier. Querying sources where they live is one response to that barrier; in the GitHub example, the agent reads audit logs through the API instead of waiting for them to reach a SIEM.

What a threat hunt produces

A hunt that finds no attacker can still succeed. PEAK's test is whether you are "satisfied that you would have found it, had it been present." The outputs worth tracking are the ones PEAK lists as deliverables and metrics:

  • Incidents. Anything suspicious goes to the SOC, triaged first so responders get a head start.
  • New and improved detections. PEAK's Hierarchy of Detection Outputs ranks them by how little human attention they need: reports, then dashboards, then analytics in code, then signatures and rules. Pick the highest level you can implement. When and how to promote a hunt is covered in when a threat hunt should become a detection.
  • Visibility gaps. A missing log source, a source that some hosts don't send, or an event type that isn't collected. MITRE's method builds this in: it runs a gap analysis of sensors and data sources before executing the hunt.
  • Vulnerabilities and misconfigurations found along the way, such as MFA not turned on.
  • Documentation and new hunt ideas, detailed enough that someone else can rerun the hunt.

Cotool's detection product is organized around these outputs. Hunts and threat intel produce detection suggestions, written as Python and backtested, that wait for a person's review. The same product maps the detection library onto MITRE ATT&CK and ranks the uncovered techniques, which turns visibility gaps into a work queue.

PEAK argues for measuring effects rather than effort: "We put twelve new automated detections into production this quarter" says more than "We performed nine hunts this quarter." In the SANS 2026 free-text responses, the most common workflow ends the same way, with a post-hunt step in which "confirmed findings are either converted into detections or assigned to platform owners for remediation."

Examples

An intel-driven hunt. A vendor publishes research on a campaign. The hunter pulls out the behaviors, checks which affected software the organization runs, and searches the last 30 days of endpoint and network data for those behaviors. Cotool runs this loop continuously: its agents crawl blogs, government databases, and corporate disclosures to produce structured intel, add the customer's own sources (existing intel vendors, STIX/TAXII feeds, or a blog), and investigate each item for relevance and exposure in the customer's environment. Real exposure or compromise produces tuned detections and next steps.

A hypothesis-driven hunt. The example on Cotool's hunt page starts from a threat model of the environment and the hypothesis "A stolen GitHub token is used to clone private repositories from an unmanaged device." The agent pulls 14 days of GitHub audit logs through the GitHub API, without centralizing them first, joins them with Okta sessions and device posture data, and finds one burst of 212 private repositories cloned in 6 minutes from an unmanaged device. It then backtests a detection for mass repository cloning, mapped to ATT&CK T1213, and sends it for review. Cotool's AI SOC frames the goal as coverage that compounds: agents hunt from fresh intel and the threat model and turn what they find into detections.

A TTP-driven hunt with stack counting. Bianco's HMM2 example is finding malware by collecting which programs are set to start automatically on every host and looking at the rarest ones. In Microsoft Defender advanced hunting, a first pass might look like this:

DeviceRegistryEvents
| where Timestamp > ago(30d)
| where RegistryKey has @"\CurrentVersion\Run"
| summarize Hosts = dcount(DeviceId), Sample = take_any(RegistryValueData) by RegistryValueName
| where Hosts <= 3
| order by Hosts asc

Entries present on only a handful of hosts are leads to investigate, not verdicts; least-frequency analysis only works when you collect from many hosts. If the hunt turns up a reliable pattern, it becomes a candidate rule, and the same evaluation applies to it as to any detection as code.

  • Detection engineering: writing, testing, and tuning the rules that run continuously. Hunts are a main source of new rules.
  • Alert triage: deciding whether a fired alert is real. See what an AI SOC is for how agents now take this first pass.
  • MITRE ATT&CK: the catalog of attacker tactics and techniques that TTP-driven hunts and coverage maps use.
  • Pyramid of Pain: a ranking of indicator types from those an attacker changes easily, such as hashes and IP addresses, up to TTPs. PEAK suggests tracking which levels your hunts cover.
  • AI threat hunting: hunts run by agents with query tools over raw telemetry. See what agents can and can't hunt yet.

FAQ

What is threat hunting in simple terms?
Threat hunting is looking for attackers in your environment on purpose, instead of waiting for a security tool to raise an alert. A hunter picks a specific question, such as whether stolen tokens are being used to clone private code, searches the logs that would show it, and reports what the data does and doesn't establish.
What are the main types of threat hunting?
The main types of threat hunting are hypothesis-driven hunts, which test an educated guess about attacker behavior; intel-driven hunts, which start from a threat report, campaign, or vulnerability; and TTP-driven hunts, which work through the techniques an attacker would need, often mapped to MITRE ATT&CK. Splunk's PEAK framework adds baseline hunts, which profile normal activity to find deviations, and model-assisted hunts that use machine learning.
What is the difference between threat hunting and threat detection?
Threat detection runs automated rules continuously and raises an alert when known patterns match, while threat hunting is a time-boxed search led by a person or agent for activity that no rule covers yet. Hunting feeds detection: once a hunt proves it can find a behavior reliably, the query usually becomes a scheduled rule so nobody has to hunt for it by hand again.
Is a threat hunt that finds nothing a failure?
A threat hunt that finds no attacker is not a failure if the hunter is confident the activity would have been found had it been present. Splunk's PEAK framework treats that outcome as a successful hunt, and the hunt still produces documentation, a candidate detection, or a list of logging gaps that kept the question from being answered.
What is the Hunting Maturity Model?
The Hunting Maturity Model, published by David Bianco in 2015, rates a threat hunting program from HMM0 (relies on automated alerting and can't hunt) to HMM4 (turns any successful hunting process into automated detection). The levels in between reflect how much data the team collects, whether it follows other people's hunt procedures, and whether it can create its own.

Sources

  1. Introducing the PEAK Threat Hunting Framework - Splunk · accessed November 16, 2026
  2. Hypothesis-Driven Hunting with the PEAK Framework - Splunk · accessed November 16, 2026
  3. Baseline Hunting with the PEAK Framework - Splunk · accessed November 16, 2026
  4. Turning Hunts Into Detections with PEAK - Splunk · accessed November 16, 2026
  5. Measuring Hunting Success with PEAK - Splunk · accessed November 16, 2026
  6. Key Threat Hunting Deliverables with PEAK - Splunk · accessed November 16, 2026
  7. A Simple Hunting Maturity Model - David J. Bianco · accessed November 16, 2026
  8. A Framework for Cyber Threat Hunting - Sqrrl · accessed November 16, 2026
  9. TTP-Based Hunting - MITRE · accessed November 16, 2026
  10. SANS 2026 Threat Hunting Survey: The Evolution of Threat Hunting · accessed November 16, 2026
  11. SANS 2026 Threat Hunting Survey: The Evolution of Threat Hunting (full report PDF) · accessed November 16, 2026
  12. Half of Threat Hunters Now Call Data Their Biggest Barrier - SANS Institute (GlobeNewswire) · accessed November 16, 2026
  13. Threat Hunting vs. Threat Detecting - Splunk · accessed November 16, 2026
  14. AI Threat Hunting and Actionable Threat Intel - Cotool · accessed November 16, 2026
  15. AI Detection Engineering with Python Detections - Cotool · accessed November 16, 2026
  16. AI SOC Without the Black Box - Cotool · accessed November 16, 2026

Keep reading

Attackers are scaling with tokens

Cotool helps defenders operate at machine speed. See how security teams are scaling Detection & Response.

Book a demo