Most hunt teams do not fail for lack of skill, they fail because every hunt is run differently, recorded nowhere and never turns into a detection. Structured hunting frameworks fix that by giving each hunt the same shape: where the idea came from, what was searched, what was concluded and what was handed over. This guide compares the three models practitioners actually use, PEAK from Splunk SURGe, TaHiTI from the Dutch financial sector and the older Sqrrl reference model (the Hunting Loop, the Hunting Maturity Model and the Hunt Matrix), and shows how to choose one, adapt it and run it as a programme. It is written for SOC leads, hunt team leads and detection engineers who already know what hunting is. For the basics of hunt types, data and identity hunting, start with the Threat Hunting Guide.
Key takeaways
- The frameworks share one loop. Every model runs trigger or idea, prepare, search, conclude and hand over; they differ in where they put the emphasis, not in the basic sequence.
- PEAK is the most complete process model. It defines three hunt types (hypothesis-driven, baseline and model-assisted) with their own Prepare, Execute and Act steps, plus the ABLE method for writing hypotheses.
- TaHiTI is the best fit when threat intelligence drives the work. It treats intelligence as the source of hunts, the context during them and a destination for their results, and it is built around a hunting backlog of abstracts.
- Sqrrl gives you the measuring stick. The Hunting Maturity Model (HMM0 to HMM4) and the Hunt Matrix describe programme maturity; they are not a step-by-step process on their own.
- Adopt one, then standardise the record. A shared hunt template, a ranked backlog and outcome metrics matter more than which framework name is on the wiki.
Why a framework, and what it actually standardises
A framework is not a methodology for finding a specific threat; it is an operating model for the hunt team. In practice it standardises five things:
- Intake: where hunt ideas come from (intelligence reports, ATT&CK gap analysis, incidents, red team findings, new data sources) and how they are written down before work starts.
- Selection: how the next hunt is picked from a backlog, so effort follows risk rather than whatever the loudest analyst read that morning.
- Execution discipline: scoping (systems, data, time window, maximum duration), analysis technique and when to refine or stop.
- Conclusion rules: what counts as proven, disproven or inconclusive, and when to escalate to incident response mid-hunt.
- Outputs: detections, data gap tickets, misconfiguration findings, intelligence updates and new backlog items, each owned by a named team.
Without those five, hunting becomes ad hoc searching. NIST recognises hunting as a distinct control in SP 800-53 Rev. 5 RA-10 (Threat Hunting), separate from continuous monitoring in SI-4, which is a useful anchor when you need to justify a dedicated process to auditors.
The three frameworks compared
| Attribute | PEAK | TaHiTI | Sqrrl reference model |
|---|---|---|---|
| Origin | Splunk SURGe, David Bianco (April 2023) | Dutch financial sector working group, lead author Rob van Os (2018) | Sqrrl, a hunting platform vendor (2015); archived by the ThreatHunting project |
| Full name | Prepare, Execute and Act with Knowledge | Targeted Hunting integrating Threat Intelligence | Hunting Loop, Hunting Maturity Model and Hunt Matrix |
| Phases | Prepare, Execute, Act; Knowledge feeds every phase | Initiate, Hunt (define/refine, execute, validate), Finalize | Hypothesis, investigate with tools and techniques, uncover patterns and TTPs, inform and enrich analytics |
| Hunt types | Hypothesis-driven, baseline (exploratory data analysis), model-assisted (M-ATH) | Hypothesis-driven, triggered by intelligence and other processes | Hypothesis-driven |
| Hypothesis method | ABLE: Actor, Behaviour, Location, Evidence | Hypothesis written during define/refine from the abstract | Educated guess from attack models, alerts, friendly and threat intelligence |
| Backlog concept | Yes: new ideas from each hunt go back to the backlog | Central: abstracts stored on a hunting backlog | Implicit (new hypotheses follow from findings) |
| Maturity model | Uses the HMM | None of its own; supported by the MaGMa for threat hunting tool | HMM0 to HMM4, Hunt Matrix |
| Metrics | Five impact measures (detections, incidents, gaps, misconfigurations, techniques) | Six measures including dwell time of findings and new intelligence created | Maturity level rather than per-hunt metrics |
| Best for | Teams wanting one model for all hunt types, including data science work | Teams with a CTI function and regulated environments that need traceable process | Assessing and explaining programme maturity |
PEAK in practice
PEAK's main contribution is that it gives each hunt type its own steps rather than forcing everything into a hypothesis. Splunk notes that steps can be skipped, reordered or added to fit the situation.
Hypothesis-driven hunts
- Prepare: pick a topic (for example data exfiltration), research it including known tools, detection gaps and published hunts, write a testable hypothesis, scope systems, data and time window with a maximum hunt duration, and plan data collection, techniques and who does what.
- Execute: gather data into one place, pre-process it (normalise formats, drop broken records), analyse with frequency analysis, clustering or visualisation, refine the hypothesis when results are empty or surprising, and escalate likely malicious activity to incident response immediately rather than at the end.
- Act: preserve the data, queries and techniques, document the result including data gaps and misconfigurations, create or hand over detections, add new ideas to the backlog and communicate to stakeholders.
ABLE: writing a hypothesis you can test
| ABLE element | Question it answers | Example (DNS tunnelling) |
|---|---|---|
| Actor (optional) | Who, if known? | An actor reported in your sector intelligence using DNS for command and control |
| Behaviour | Which one or two techniques? | T1071.004 Application Layer Protocol: DNS |
| Location | Where would it appear? | Finance department workstations resolving through internal DNS servers |
| Evidence | Which data, and what would it look like? | DNS query logs: unusually long query names, high volumes of TXT records, many unique subdomains for one parent domain |
The discipline is narrowing: topic to technique to a specific population and data source. A hypothesis such as "attackers may be exfiltrating data" is a topic, not something you can prove or disprove.
Baseline and model-assisted hunts
- Baseline (exploratory data analysis): choose one data source, group similar systems (desktops, application servers) and baseline each group separately over roughly 30 to 90 days. Build a data dictionary, review distributions and cardinality, investigate outliers with stack counting, z-scores or isolation forests, and record known-benign outliers. PEAK's own caution applies: abnormal does not mean malicious, so many baselines become dashboards rather than alerts.
- Model-assisted (M-ATH): uses machine learning models of known good or known malicious behaviour to surface candidates for a hunter to review. It suits teams with data science support and stable, well-parsed data; without those, start with baselines.
TaHiTI in practice
TaHiTI was written because its authors found few published methodologies that covered the whole hunting process. Its distinctive features are the abstract, the backlog and explicit handover.
- Initiate: a trigger arrives (a threat intelligence report on an actor relevant to your sector and region, an incident, a red team result, a new ATT&CK technique) and is written up as a short hunt abstract, then stored on the hunting backlog. The paper suggests something as simple as SharePoint or JIRA; the point is that the team can see enough to choose the most relevant abstract next.
- Hunt, define/refine: the abstract becomes a full investigation plan with data sources, analysis techniques and a hypothesis.
- Hunt, execute: data is retrieved and analysed, with iterative refinement.
- Hunt, validate: the hypothesis is proven, disproven or inconclusive. An inconclusive result loops back to define/refine instead of closing silently.
- Finalize: results, conclusions, recommendations (preventive measures, logging, monitoring use cases, process) and lessons learned are documented, then handed to incident response, security monitoring, threat intelligence and vulnerability management.
Threat intelligence plays three roles in TaHiTI: a starting point for hunts, context during the hunt (pivoting to related techniques with ATT&CK and the diamond model) and a consumer of hunt output, such as newly observed techniques that feed actor profiles and peer sharing. That third role is what most programmes miss. See the Threat Intelligence Guide and the Threat Profiling for Your Organisation Guide.
The Sqrrl reference model: loop, maturity and matrix
Sqrrl operated from about 2015 until Amazon Web Services acquired it in early 2018, and its site went offline in 2019. The ThreatHunting project now hosts the key papers with Sqrrl's permission. The model has three parts.
- The Hunting Loop (October 2015): form a hypothesis about activity that may be happening, informed by attack models, existing alerts and friendly and threat intelligence; investigate it with tools and techniques such as linked data search, visualisation, statistics and machine learning; uncover new patterns and TTPs, which Sqrrl called the definitive success criterion; and inform automated analytics so successful techniques stop needing a human.
- The Hunting Maturity Model: described by David Bianco in 2015, it is the most cited part and PEAK still uses it.
- The Hunt Matrix: crosses the loop with the maturity levels, showing how each stage improves: data collection grows from little to many data types, hypotheses draw on more intelligence, procedures and tools become richer, pattern detection moves up the Pyramid of Pain and findings feed automation more routinely.
| Level | Name | What distinguishes it | Typical next step |
|---|---|---|---|
| HMM0 | Initial | Relies on automated alerts; little data collected beyond what tools need. Not capable of hunting. | Centralise endpoint, identity and DNS logs with searchable history |
| HMM1 | Minimal | Routinely collects some data; searches history for indicators from intelligence reports | Move from indicator sweeps to behaviour-based procedures |
| HMM2 | Procedural | Runs procedures written by others, mostly least-frequency analysis; the most common level among active programmes | Train hunters to write and publish their own procedures |
| HMM3 | Innovative | Several hunters create and document their own repeatable procedures; scheduling becomes a scaling problem | Build a pipeline from hunt to production detection |
| HMM4 | Leading | Successful hunts are routinely automated into detection, freeing hunters for new work | Measure technique coverage and retire stale hunts |
Choosing and combining frameworks
- Small team or SOC rotation: use PEAK hypothesis hunts with ABLE and a one-page template. Skip M-ATH until data quality is good. Use the HMM once a year to show progress.
- Organisation with a CTI function: use TaHiTI's Initiate phase and backlog for intake, because it makes the intelligence-to-hunt link explicit, and PEAK's Execute and Act steps for the work itself. Many mature teams run exactly this hybrid.
- Data-rich environment with data science support: PEAK, because it is the only one of the three that treats baselines and model-assisted hunting as first-class hunt types.
- Regulated financial or threat-led testing context: TaHiTI's documentation and handover steps produce the audit trail supervisors expect. See the Threat-Led Penetration Testing Guide.
- Board or audit reporting: report the HMM level and the outcome metrics, not the framework phases.
Whatever you choose, ATT&CK is the common vocabulary. MITRE's TTP-based hunting paper (2020) argues the same point from first principles: adversaries are constrained by the technology they operate on, so hunting by technique is more durable than hunting by indicator. Use the Threat-Informed Defence Guide to pick techniques.
A standard hunt record
The record is where frameworks pay off. A template that merges TaHiTI's abstract with PEAK's documentation steps works for all three models:
| Field | Filled at | Content |
|---|---|---|
| Hunt ID, title, owner | Intake | Unique ID, short name, accountable hunter |
| Trigger and source | Intake | Intelligence report, incident, red team finding, coverage gap, new data source |
| Abstract and priority | Intake | Two or three sentences; risk-based priority score |
| Hypothesis (ABLE) | Prepare | Actor, behaviour with ATT&CK ID, location, evidence |
| Scope | Prepare | Systems, data sources, time window, maximum duration |
| Queries and notebooks | Execute | Version-controlled link, not pasted text |
| Result | Validate | Proven, disproven or inconclusive, with evidence |
| Escalations | Execute | Incident ticket numbers and times |
| Outputs | Act / Finalize | Detections created or updated, data gaps, misconfigurations, intelligence updates, new backlog items, each with an owner ticket |
| Lessons learned | Act / Finalize | What would make the next hunt faster |
Community libraries help fill the backlog. HEARTH, maintained by the THOR Collective, publishes several hundred hypotheses categorised by PEAK type (Flames for hypothesis-driven, Embers for baselines, Alchemy for model-assisted). The OTRF ThreatHunter-Playbook publishes hunts as Jupyter notebooks organised by ATT&CK tactic with recorded datasets for practice. Treat both as idea sources to adapt, not as hunts to run unchanged.
Rolling out a framework-based programme
- Confirm data readiness first: list the data sources each candidate hunt needs and confirm 30 to 90 days of searchable history. See the Log Source Coverage Matrix Template.
- Pick the framework and adapt it in writing: one page saying which phases you use, your conclusion rules and your escalation threshold.
- Stand up the backlog and template: seed 15 to 20 abstracts from your threat profile, recent incidents and coverage gaps; rank by risk.
- Pilot for one quarter: time-box each hunt (for example two weeks), run four to six hunts, and review each record with a detection engineer.
- Build the handover path: agree with detection engineering how a hunt query becomes a rule, tested in audit or report-only mode before it alerts. See the Detection Engineering Guide.
- Measure and adjust: after the pilot, check outcome metrics, retire template fields nobody uses and assess the HMM level.
- Roll back gracefully: if the process slows hunting more than it helps, cut the template to the mandatory fields (hypothesis, scope, result, outputs) rather than abandoning records altogether.
Detection and measurement
PEAK and TaHiTI agree on measuring effects rather than activity. Combined, the useful metrics are:
- Detections created or updated per quarter from hunts, including false positive reductions to existing rules.
- Data gaps identified and closed: missing sources, parsing faults, short retention, with closure rate.
- Vulnerabilities and misconfigurations found and remediated as a result of hunts.
- Unique ATT&CK techniques hunted against your prioritised technique list.
- Dwell time of findings: how long malicious activity had been present when a hunt found it (TaHiTI).
- Backlog health: age of the oldest high-priority abstract and share of hunts that came from intelligence.
PEAK notes that a raw count of incidents opened during hunts is a weak indicator; do not score hunters on it, or they will hunt where they already expect noise.
Where identity fits
Identity is where hypothesis-driven hunting pays off quickest, because the evidence (sign-in, token and audit logs) is centralised and the behaviours are well described. ABLE examples include service accounts signing in interactively (T1078 Valid Accounts), new credentials added to application registrations, and API keys used from new networks. Non-human identities rarely appear in hunting backlogs because no one owns them. See the Identity Threat Hunting Guide and the Detection Engineering for Non-Human Identities Guide.
Framework mapping
| Framework | Reference | Relevance |
|---|---|---|
| NIST SP 800-53 Rev. 5 | RA-10 Threat Hunting | Establishes a hunting capability; the frameworks describe how to run it |
| NIST SP 800-53 Rev. 5 | SI-4 System Monitoring, IR-4 Incident Handling, PM-16 Threat Awareness Program | Detections from hunts, escalation, intelligence sharing |
| NIST CSF 2.0 | DE.AE-02, DE.AE-07, DE.CM-01, ID.RA-03 | Analysis of adverse events, integrating intelligence, monitoring, recording threats |
| ISO/IEC 27001:2022 Annex A | A.5.7 Threat intelligence, A.8.16 Monitoring activities | Intelligence input and monitoring improvements from hunts |
| CIS Controls v8.1 | Control 8 Audit Log Management, Control 13 Network Monitoring and Defense | Data that hunting depends on |
| MITRE ATT&CK v20 | Technique IDs in hypotheses, for example T1071.004, T1048, T1059.001 | Common vocabulary for behaviour and coverage |
Common failure patterns
- Framework on the wiki, not in the work: a process diagram is adopted but hunts still have no record, no conclusion and no outputs.
- Topics masquerading as hypotheses: "look for ransomware" cannot be disproven. Force every hunt through ABLE or an equivalent before it starts.
- Indicator sweeps counted as hunts: searching last month's hashes is HMM1 work. It has value, but report it separately so maturity is not overstated.
- No maximum duration: hunts drift for weeks because nothing says when to conclude inconclusive and move on.
- Findings that never leave the hunt team: no agreed path to detection engineering, data owners or intelligence, so the same gaps are rediscovered every quarter.
- Escalating only at the end: PEAK is explicit that likely malicious activity goes to incident response immediately, not in the final report.
Practitioner checklist
- Choose PEAK, TaHiTI or a documented hybrid, and write a one-page adaptation.
- Adopt a single hunt record template with hypothesis, scope, result and outputs as mandatory fields.
- Require ABLE (or equivalent) hypotheses with an ATT&CK technique ID.
- Run a ranked hunting backlog fed by intelligence, incidents, red team findings and coverage gaps.
- Set a maximum duration per hunt and a rule for inconclusive results.
- Agree escalation criteria with incident response before the first hunt.
- Version-control hunt queries and notebooks; link them from the record.
- Agree a handover path to detection engineering with audit-mode testing.
- Report outcome metrics quarterly and assess the HMM level annually.
Standards and references
- Splunk SURGe: Introducing the PEAK Threat Hunting Framework
- Splunk SURGe: PEAK Hypothesis-Driven Threat Hunting
- Splunk SURGe: PEAK Threat Hunting Metrics
- Dutch Payments Association: TaHiTI Threat Hunting Methodology
- ThreatHunting Project: Sqrrl Archive (Hunting Loop and Hunt Matrix)
- David J. Bianco: A Simple Hunting Maturity Model
- NIST: SP 800-53 Rev. 5
- MITRE: TTP-Based Hunting
Related NHI Mgmt Group resources: Threat Hunting Guide · KQL for Threat Hunting Guide · Detection Engineering Maturity Model · Correlation Rules and Risk Based Alerting Guide