Skip to main content
Dithered dot illustration of a mechanical stopwatch, with dots scattering off one edge

SOC metrics: MTTD, MTTR, and what to measure instead

How to define and measure MTTD, MTTA, MTTR, and mean time to contain from SIEM timestamps, which benchmarks to trust, and what changes when agents triage.

Metrics
Updated 11 min read

Summary

MTTD (mean time to detect) measures how long malicious activity runs before the SOC raises an alert; MTTA measures how long an alert waits for someone to pick it up; MTTR measures how long it takes to resolve an incident once detected, and mean time to contain covers the part of that before the threat stops spreading. Measure each from SIEM and case timestamps, report medians and 90th percentiles by severity, and keep false positives separate. Published benchmarks range from a 3-day median dwell time (Sophos) to a 247-day mean breach lifecycle (IBM) because they measure different populations.

Global median dwell time, Mandiant investigations in 2025 (M-Trends 2026)
14 days
Median dwell time across 661 Sophos IR and MDR cases
3 days
Mean time to identify plus contain a breach, IBM 2026
183 + 64 days
SOCs citing incidents handled as a top metric (SANS 2026)
70%

Key takeaways

  1. Write down the start and end timestamp for every metric before you report it.
  2. Report medians and 90th percentiles by severity; a few long cases distort a mean.
  3. Split true positives from false positives, or fast false-positive closures will flatter MTTR.
  4. Compare against your own history; public benchmarks measure different populations.
  5. When agents triage every alert, track escalation quality and coverage, not just speed.

Formulas

Mean time to detect (MTTD)

MTTD = sum of (alert or incident created − first malicious activity) / number of true-positive incidents

End time is the alert or incident creation time. Start time is the earliest attacker activity found during investigation; the SIEM's first event time is only a lower bound.

Mean time to acknowledge (MTTA)

MTTA = sum of (first human or agent action − alert created) / number of alerts

In Microsoft Sentinel, FirstModifiedTime − CreatedTime. Exclude modifications by automation rules, or acknowledgment looks instant. In Cotool, the first agent execution or user action on the alert, which the audit log records separately.

Mean time to investigate

Mean time to investigate = sum of (disposition recorded − alert created) / number of alerts

Disposition is the closing classification (true positive, benign, false positive) or the escalation decision, taken from the case system. In Cotool, the activity-timeline entry where the alert moves to a Closed status or is escalated.

Mean time to contain (MTTC)

MTTC = sum of (containment complete − incident created) / number of true-positive incidents

SIEMs rarely timestamp containment. Record it as a case field or task, such as host isolated or account disabled. In Cotool, the audit-log entry for the approved containment tool call, such as a host isolation.

Mean time to respond or resolve (MTTR)

MTTR = sum of (incident closed − incident created) / number of incidents

In Microsoft Sentinel, ClosedTime − CreatedTime. ClosedTime is the last close, so reopened incidents carry their full duration.

False-positive rate

False-positive rate = alerts closed as false positive / alerts closed

Comes from the closing classification. Track it per detection rule; a rule-level rate is what tells you what to tune.

Alert-to-incident ratio

Alert-to-incident ratio = alerts generated / incidents opened

In Sentinel, count AlertIds per incident. A high ratio can mean good grouping or a noisy rule; read it alongside the false-positive rate.

Detection coverage

Coverage = ATT&CK techniques with a tested detection / techniques in your threat model

Count a technique only when a detection has been tested against real or replayed data, not when a rule merely exists.

SOC time metrics look simple and are easy to get wrong. Two teams can both report an MTTR of four hours and mean entirely different things. This guide defines each metric with its start and end timestamps, shows where those timestamps come from, and explains which public benchmarks are worth citing.

What each metric measures

The metrics follow the life of a threat from first activity to closure:

MetricStarts whenEnds when
Mean time to detect (MTTD)Attacker activity beginsAlert or incident is created
Mean time to acknowledge (MTTA)Alert is createdSomeone, or an agent, picks it up
Mean time to investigateAlert is createdA disposition is recorded
Mean time to contain (MTTC)Incident is createdThe threat can no longer spread
Mean time to respond or resolve (MTTR)Incident is createdIncident is closed

MTTD measures your detection engineering and telemetry. MTTA and mean time to investigate measure the alert queue. MTTC and MTTR measure response. Watch the acronyms: IBM uses MTTI for mean time to identify a breach, which is closer to MTTD, while many SOC tools use MTTI for mean time to investigate.

Benchmarks

Most widely quoted SOC benchmarks come from three studies, and they disagree by two orders of magnitude because they measure different populations.

StudyFigurePopulation and method
IBM Cost of a Data Breach 2026Mean 183 days to identify, 64 to contain, 247 total602 breached organizations, March 2025 to February 2026; interviews
Mandiant M-Trends 2026Median dwell time 14 days, up from 11Mandiant Consulting investigations of targeted attacks in 2025
Sophos Active Adversary Report 2026Median dwell time 3 days661 Sophos IR and MDR cases, November 2024 to October 2025

IBM's report, conducted by the Ponemon Institute, covers data breaches of 2,590 to 115,380 records across 17 industries and 16 countries, based on interviews with 3,558 people. It reports means, so long breaches pull the figures up. It also splits the lifecycle by who found the breach: 209 days when internal teams identified it, 230 days for MSSPs, and 281 days when a third party did. Organizations using security AI and automation extensively averaged 215 days, against 280 days for those with no use; that's a correlation, not a controlled test.

M-Trends 2026 reports dwell time, how long attackers stayed on a network undetected, as a median. Mandiant attributes the rise to 14 days largely to espionage and North Korean IT worker cases, which had a median of 122 days. Organizations found the intrusion themselves 52% of the time, up from 43% in 2024.

Sophos reports a median of 3 days, with 5 days for incident response cases and 2 days for managed detection and response (MDR) customers. 69% of its cases come from MDR customers, who are already monitored, which pulls the overall median down.

For alert-level metrics such as MTTA, investigation time, and false-positive rate, we found no public benchmark that publishes its sample and method. Vendor figures without methodology can't be compared with your numbers, so baseline against your own history instead.

How to measure

Every metric needs two timestamps from systems you control. In Microsoft Sentinel, the SecurityIncident table holds most of them:

  • FirstActivityTime: the first event in the incident's alerts, the earliest detection-side evidence.
  • CreatedTime: when the incident was created.
  • FirstModifiedTime: when the incident was first modified, Microsoft's basis for time to triage.
  • ClosedTime: when the incident was last closed.
  • Classification, ModifiedBy, AlertIds, and Tasks: the closing classification, who made each change, the grouped alerts, and the tasks on the incident.

The table leaves gaps you have to fill. It has no containment column, so MTTC needs a task or a custom field. FirstActivityTime is the first event in the incident's alerts, not the first attacker activity an investigation later finds. FirstModifiedTime records the first change from any source, automation rules included. And it holds Sentinel incidents only, so alerts triaged only in another tool's console never reach it.

Every incident update writes a new row, so keep only the latest row per incident with arg_max, as Microsoft's sample queries do; those queries report triage and closure times as 5th, 50th, 90th, and 99th percentiles. A starting query:

SecurityIncident
| summarize arg_max(TimeGenerated, *) by IncidentNumber
| where Status == "Closed"
| extend DetectHours = (CreatedTime - FirstActivityTime) / 1h,
         TriageHours = (FirstModifiedTime - CreatedTime) / 1h,
         CloseHours = (ClosedTime - CreatedTime) / 1h
| summarize p50_detect = percentile(DetectHours, 50),
            p50_triage = percentile(TriageHours, 50),
            p50_close = percentile(CloseHours, 50),
            p90_close = percentile(CloseHours, 90)
            by Severity, Classification

Other SIEMs and case tools have equivalents: an alert creation time, a first-action time, a closure time, and a classification. Containment is the timestamp most teams lack. Add it as a required case field or task so MTTC comes from data, not memory.

An agent-run queue records the same points in different places. In Cotool, for example, the alert's activity timeline records every status, severity, and assignee change, assigning a response agent starts its triage run, and alerts close as True Positive, False Positive, or Benign, which gives you the classification split (Cotool docs). Containment comes from the audit log: an approved host isolation is a tool call, logged with its parameters alongside every agent execution and user action, filterable by date range and exportable by API to your SIEM (Cotool audit logs), so MTTC can end at a recorded action instead of a note.

Pitfalls that distort the numbers

Undefined endpoints. "Detect" can start at first attacker activity, first logged event, or first alert, and each gives a different MTTD. The SIEM only knows what it saw; true first activity often comes from the investigation, so update it on the case when you find it.

Averages instead of medians. Dwell time and resolution time have long tails. M-Trends cites intrusions involving the BRICKSTORM backdoor that averaged 393 days of dwell time; one such case can move a quarterly mean more than every other incident combined. Report the median and 90th percentile.

Mixing true and false positives. False positives close quickly, so including them lowers MTTR without anything improving. Report MTTR and MTTC on true positives, report false-positive rate separately, and use mean time to investigate for all alerts.

Automation timestamps. If an automation rule assigns or tags incidents on creation, FirstModifiedTime becomes the automation's timestamp and MTTA looks instant. Filter on ModifiedBy, or measure first human action from the case history. The same split applies when agents triage: Cotool's audit log can be filtered by user, by agent, and by action type, so agent pickup and first human action come out as separate timestamps (Cotool audit logs).

Counting volume as performance. The 2026 SANS SOC Survey found that 70% of respondents cite the number of incidents handled as a top metric, which it has been for 10 consecutive years, and that it's also the top metric used to set SOC budget. SANS argues that incident counts measure how busy the SOC is, not whether it caught the right threats early, and recommends time from detection to containment instead.

How AI agents change the numbers

When agents investigate every alert, several metrics change for definitional reasons, not just because work got faster.

  • MTTA approaches agent pickup time. The human queue moves to escalations, so measure time from escalation to human decision as its own metric.
  • The denominator changes. If analysts previously skipped or bulk-closed low-severity alerts, false-positive rate and investigation time only described the alerts people touched. With every alert dispositioned, those rates cover everything, and they can look worse before they look better. Note the change point on your charts.
  • New metrics matter more. Track escalation rate (alerts escalated divided by alerts dispositioned), agreement rate (of a sample of agent dispositions, the share a human reviewer upholds), and how many detections were tuned as a result of dispositions.
  • MTTD doesn't fall on its own. Faster triage only helps once an alert fires. M-Trends found the median time from initial access to hand-off to a second threat group fell to 22 seconds, and Sophos found attackers reached Active Directory a median of 3.4 hours after initial access. Detection still depends on collecting and keeping the right telemetry, and on proactive hunting for activity no rule covers yet; Sophos found retention-related missing logs doubled year over year, driven largely by firewalls with short default retention.

Those new metrics need inputs. In Cotool, for example, every verdict shows the evidence and reasoning behind it, and an analyst can reopen any investigation and ask follow-up questions, so a sampled agreement review checks the agent's work rather than redoing it. An AI judge grades every run against acceptance criteria you set and flags a broken integration when it fails, which catches an agent that has stopped querying the EDR before it shows up as a shift in escalation rate. Its 0-100 score per run is recorded in run history (Cotool docs), so you can weight the agreement sample: review every run scored below 60 plus a random slice of the rest, and report agreement for each group separately. For detections that run in Cotool, each false-positive or benign disposition feeds autotuning, which tests a change against the current version before applying it or sending it for review.

Adoption is broad but shallow. IBM found 50% of breached organizations deploying AI agents in their SOC, and SANS found that 79% of SOCs use AI or ML tools but only 36% have built them into a defined workflow. For background, see what an AI SOC is and what alert fatigue does to a queue.

When agents take the first response

Cotool's agents pick up every alert as it lands, gather context across your tools, and close or escalate it with the evidence attached. State-changing actions such as isolating a host can require approval in the app or Slack, so containment stays a human decision you record as its own step. In that model, MTTA and investigation time mostly measure the agents' configuration and the tools they can reach, and the metrics a team works on directly are escalation quality, detection coverage, and time from escalation to containment.

FAQ

What is MTTD and MTTR in a SOC?
MTTD, mean time to detect, is the average time between the start of malicious activity and the SOC raising an alert or incident. MTTR, mean time to respond or resolve, is the average time from that alert or incident to closing it. Teams define both endpoints differently, so check the definition before comparing an MTTD or MTTR figure with anyone else's.
What is a good MTTR for a SOC?
No public benchmark gives a good MTTR for a SOC that you can apply directly, because published studies measure different things. IBM reports a mean of 64 days to contain a data breach, while Microsoft Sentinel's own metrics measure closure in hours from incident creation. Set MTTR targets by severity from your own median and 90th percentile, then improve against them.
How do you reduce MTTD and MTTR?
You reduce MTTD by collecting the telemetry attackers actually touch, keeping it long enough to investigate, and testing detections against it. You reduce MTTR by cutting the time alerts wait in a queue, which is where automated or agent-led triage helps most, and by pre-approving containment actions. Measure both by severity so a flood of fast low-severity closures doesn't hide slow critical incidents.
Should false positives be included in MTTR?
False positives should be reported separately from MTTR for true positives. False positives usually close fast, so mixing them in pulls MTTR down and hides how long real incidents take. Track the false-positive rate as its own metric and use mean time to investigate for all alerts.
What KPIs should a SOC track besides MTTD and MTTR?
Besides MTTD and MTTR, a SOC should track mean time to acknowledge, mean time to contain, false-positive rate per detection, alert-to-incident ratio, and detection coverage against its threat model. The 2026 SANS SOC Survey found that the number of incidents handled has been the top reported SOC metric for 10 years, and argues it measures volume rather than value.

Sources

  1. Cost of a Data Breach Report 2026 (IBM, Ponemon Institute) · accessed November 16, 2026
  2. Cost of a Data Breach Report 2026 | IBM · accessed November 16, 2026
  3. M-Trends 2026 Report: Executive Edition · accessed November 16, 2026
  4. M-Trends 2026: Data, Insights, and Strategies From the Frontlines | Google Cloud Blog · accessed November 16, 2026
  5. Nowhere, man: The 2026 Active Adversary Report | Sophos · accessed November 16, 2026
  6. 2026 SANS SOC Survey Insights: A Decade of Evolution in Cyber Defense (SANS Institute) · accessed November 16, 2026
  7. Use incident metrics in Microsoft Sentinel to measure SOC performance · accessed November 16, 2026
  8. Azure Monitor Logs reference - SecurityIncident · accessed November 16, 2026
  9. AI SOC Without the Black Box · accessed November 16, 2026
  10. Working Alerts - Cotool Documentation · accessed November 16, 2026
  11. Audit Logs - Cotool Documentation · accessed November 16, 2026
  12. Automatic Evaluations - Cotool Documentation · 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