top of page
40051abd5a76713af8f015988fc6780e-blue-phone-icon-with-a-wave-on-it.webp
Logo-Zalo-Arc.webp
Facebook_Messenger_logo_2020.svg.webp
whatsapp

Threat hunting in the SOC: from hypothesis to evidence and new detections

2 minutes ago
9 min read

Most SOCs are alert-driven: when a rule fires, an analyst investigates it. The blind spot is behavior that no rule covers yet, especially when attackers use valid accounts and administration tools already present in the environment. According to the SANS 2026 Threat Hunting Survey, 73% of nation-state attacks still rely on living-off-the-land techniques, yet only 37% of organizations have a formally defined hunting methodology. This guide explains how a Security Operations Center (SOC) can run hypothesis-driven threat hunting, from the reason to start a hunt to evidence and new detections.

Quick answer: Threat hunting in the SOC is the practice of proactively testing a hypothesis about attacker behavior that existing alerts have not detected. A structured hunt follows 7 steps: select a trigger, write a testable hypothesis, validate the data, query and analyze, reach findings and hand off, turn the results into new detections, then measure with KPIs. The output of a hunt is evidence and detections, not just a report.
threat-hunting-in-soc-lifecycle
Threat Hunting in SOC lifecycle

Steps 1–2: From trigger to a testable hypothesis

A hunt starts when there is a specific reason for suspicion, not when an analyst has spare time and opens the SIEM. The boundary with day-to-day operations is clear: the SOC alert triage process handles what rules have already caught, while threat hunting looks for what rules have not caught yet.

Common triggers

  • Threat intelligence: reports on campaigns, threat actors or TTPs targeting the organization’s industry. The TaHiTI methodology is built around using threat intelligence as input for hypotheses.

  • Detection gaps: an ATT&CK technique that matters to the organization but is not covered by any rule.

  • Previous incidents or near-misses: check whether similar techniques still exist on other systems.

  • Infrastructure changes: new systems moving to the cloud, new partners granted access, new applications exposed to the internet.

  • Baseline anomalies: a metric that has increased without explanation.

IPSIP’s Threat Hunting Report 2025 is also a source of triggers: each trend in the report can become a topic to re-examine in the organization’s own environment.

Writing a testable hypothesis

A hypothesis is a statement that data can prove true or false. Splunk’s PEAK framework uses the ABLE model to break a hypothesis into four parts: Actor (who?), Behavior (what?), Location (where?) and Evidence (what are the expected traces?).

Running example used throughout this guide:

An attacker with valid credentials is using built-in remote administration tools to move laterally between servers outside business hours (ATT&CK T1078 Valid Accounts, T1021 Remote Services). If this is true, the SOC will see remote logon sessions from workstations to servers outside normal hours, from accounts that rarely perform this activity.

Note: the examples, queries and templates in this guide are suggestions for reference. Organizations should adapt them to their own infrastructure, data sources, operating processes and actual risk level.

Weak hypotheses tend to be too broad, such as “an intruder may already be inside,” or fail to state the expected evidence, so they cannot lead to a conclusion.

Hunt card: a one-page brief before you start

  • The trigger and where it came from

  • The hypothesis in ABLE format and related ATT&CK techniques

  • Scope: systems, user groups, data time range

  • Data sources you plan to use

  • Time-box and stopping criteria

  • Who runs the hunt and who receives the results

This card makes hunts repeatable and feeds the KPIs in step 7.

Step 3: Data requirements, validating data before the hunt

Without the right data, a hypothesis cannot be refuted, and a “nothing found” conclusion only creates a false sense of security. According to the SANS 2026 Threat Hunting Survey, 50% of respondents named data quality or volume as their biggest barrier, overtaking the shortage of skilled staff for the first time; 32% said cloud infrastructure is the hardest environment to hunt in. In Vietnam, Cisco’s Relentless Defense 2026 report found that 45% of professionals spend more time collecting and reconciling data across systems than responding to actual threats.

Mapping hypotheses to log sources

Since v18 (October 2025), MITRE ATT&CK has replaced the detection descriptions in each technique with Detection Strategies and Analytics, which point to Data Components and specific log sources. This is a good starting point for building the data list for each hypothesis instead of guessing.

For the lateral movement example using valid accounts, the SOC typically needs:

  • Logon logs on servers, such as Windows event 4624 with a remote logon type, to see who logged on where and when.

  • Process telemetry from EDR, to see which administration tools ran and their parent process.

  • Authentication logs from Active Directory or the identity provider, to correlate accounts with source devices.

  • Internal network flow data from NDR or internal firewalls, to see RDP, SMB and WinRM connections between workstations and servers.

Five data readiness questions

  • Coverage: do logs come from every system in the hunt scope, or only some?

  • Depth: do logs include the necessary fields, such as user, source host, process and command line?

  • Retention: does the data go back far enough to cover the hypothesis time window? Weigh the costs in our guide on how long to retain logs for the SOC

  • Latency: how long do logs take to reach the SIEM, and are there gaps caused by agents losing connectivity?

  • Normalization: are usernames, hostnames and time zones consistent across sources so they can be joined?

If any answer is “no,” record that gap as a hunt finding and turn it into a request for additional logs. This is still a valuable output, because it shows where the SOC cannot see.

Step 4: Query and analysis, proving or refuting the hypothesis

The goal of this step is to reach an evidence-based conclusion, not to find something suspicious at any cost. The effective approach is to go from broad to narrow: the first query identifies candidates, and later steps rule out legitimate activity.

Three analysis techniques

  • Compare against the baseline: does this account normally log on to this server, is it in an admin group, is there a change ticket or maintenance window that matches the timing?

  • Pivot on entities: from a suspicious account, look at the other machines it touched, the processes run after logon and the resulting network connections. Telemetry from EDR and XDR in the SOC is usually the fastest place to pivot.

  • Prioritize behavior over indicators: malicious hashes or IPs are easy to change, while behavior chains are much harder to change. This is the idea behind the Pyramid of Pain, which both PEAK and TaHiTI use.

During analysis, record every query run, when it ran, the number of results and why each group was ruled out. These notes become the evidence for your findings and the raw material for writing detections in step 6.

Step 5: Findings and handoff, getting results to the right owner

Every hunt must end with one of four conclusions, and each conclusion has its own recipient. A hunt that finds no malicious activity is not a failure, as long as the data is sufficient to refute the hypothesis.

Conclusion

How to recognize it

Action

Recipient

Required output

Confirmed malicious

Evidence shows unauthorized activity is happening or has happened

Escalate immediately, do not wait for the time-box to end

Incident response team (IR/CSIRT)

Timeline, affected accounts and hosts, queries run, estimated severity

Legitimate but risky

Permitted behavior that bypasses controls, for example an administrator using direct RDP instead of the privileged access management system

Do not open an incident, recommend tighter controls

System owner, IT team

Specific recommendations, list of confirmed exceptions

Refuted, sufficient data

No expected evidence within the scope and time range checked

Close the hunt, consider continuous monitoring

Detection engineer

Completed hunt card, scope checked, proposed rule

Inconclusive

Missing logs, logs that cannot be correlated, or time-box exceeded

Record the gap, reschedule the hunt

SOC lead, log platform operations team

Request for additional data, rescheduled hunt date

The matrix above is a suggested framework. Recipients and outputs should be adapted to each organization’s structure and service contracts.

When handing off to incident response

Hunters do not handle incidents themselves. Once there is evidence of malicious activity, the hunter’s job is to hand over enough for IR not to restart the investigation from scratch: the affected entities, the earliest time frame found, queries that can be rerun and what has not been checked yet. The boundaries of responsibility between teams are explained in how SOC, CSIRT and CERT differ.

Step 6: Detection feedback, turning hunts into new detections

A hunt only creates lasting value when what was found manually is detected automatically next time. The PEAK framework dedicates an entire phase (Act) to this: documenting, automating and sharing results, even when the hunt finds nothing.

Four types of output for the detection team

  • New rules or analytics: built from refined hunt queries.

  • Tuning existing rules: when a hunt shows an existing rule is too noisy or misses variants.

  • Context data: lists of admin accounts, sensitive servers and maintenance windows that help rules separate legitimate from anomalous activity.

  • Data requests: additional log sources needed for a rule to work.

From hunt query to production rule

  1. Rewrite the logic from the hunt query, preferably in a format that is portable across platforms, such as Sigma, if the SOC uses multiple tools.

  2. Test it against historical data to estimate alert volume and false positive rate.

  3. Add only the exceptions confirmed in step 5, not exceptions based on assumptions.

  4. Assign ATT&CK techniques, severity and short triage guidance for on-shift analysts.

  5. Deploy to production and monitor it for a period before considering it complete.

For this guide’s example, a reasonable output is a rule that alerts when an account outside the admin list opens a remote logon session to a server outside business hours, excluding registered maintenance windows.

Once the rule runs reliably, update the coverage matrix to track which techniques are covered. 

Step 7: KPIs to measure threat hunting effectiveness

If you cannot measure it, it is hard to protect the hunting budget. According to the SANS 2026 Threat Hunting Survey, only 40% of organizations formally measure hunting effectiveness, down from 64% in 2024. In Vietnam, Cisco’s Relentless Defense 2026 found that 35% of organizations consider their cybersecurity metrics too technical to persuade leadership.

The most misleading KPI is “number of threats found”: it depends on whether an attacker is present at all, and it pushes hunters to look for anything suspicious rather than test hypotheses. Measure the outcomes of the process instead:

  • Hunts completed versus plan for the period, with the share that reached a clear conclusion.

  • New or improved detections produced by hunts.

  • ATT&CK techniques validated in the live environment.

  • Data gaps identified and closed.

  • Time from trigger to conclusion, to see whether the process gets stuck at the data or analysis stage.

  • Hunts escalated to incident response, with the severity of each incident.

When reporting to leadership, translate these metrics into risk language.

In-house or SOC/MDR with threat hunting

An organization can hunt in-house when it meets three conditions: its data passes the five questions in step 3, it has people with sufficient analytical skills, and it has time dedicated to hunting rather than only hunting when the alert queue is empty. Without any one of the three, hunting tends to become ad hoc. According to SANS 2026, 58% of organizations still keep hunting fully in-house, while fully outsourced hunting has fallen to 18%.

Many organizations choose a middle path: the internal team holds the business context and makes decisions, while an external partner handles monitoring and periodic hunts. When comparing service types, the difference in scope between SOC and MDR often determines whether hunting is included in the contract. If budget is the main concern, you can also review the cost factors of outsourcing a SOC.

Whether you run the SOC in-house or work with a partner, the 7-step lifecycle above can serve as a yardstick for a hunting program. Do hunts start from a clear trigger and hypothesis, or just sweep IoCs from a list? Is the data validated before concluding “nothing found”? Does each hunt end with a specific conclusion and a handoff to the right owner, as in the matrix in step 5? Do the results become new detections and get reported through the KPIs in step 7? Wherever the answer is vague is where to focus improvement first.

Build the monitoring foundation for threat hunting with IPSIP Vietnam

Threat hunting only produces results when the SOC has sufficiently deep data, alerts are validated promptly and handoff processes are clear. Organizations reviewing their log sources across SIEM and EDR/XDR, or wanting to know whether their current monitoring is ready for hunting, can explore IPSIP Vietnam’s 24/7 SOC service. IPSIP provides continuous cybersecurity monitoring, combining technology with a team of specialists to validate alerts, investigate and coordinate incident response within the scope of the service.

Organizations can book a consultation with IPSIP to discuss the current state of their monitoring data, how to divide responsibilities between internal teams and partners, and a deployment approach that fits their risk profile.

ipsip-vietnam

References:

follow ipsip vietnam.png
bottom of page