Skip to content
How detection works

From the first signal to the action that ends it.

The agent watches credential theft, persistence, defence evasion, lateral movement and pre-ransomware behaviour in real time. What reaches an analyst is not that stream. It is a single case that asks the only question worth asking first: is this real?

app.secuno.nl/investigations/INV-84
Investigation INV-84 in Secuno, a critical real-time detection raised by the agent's urgent uplink: case age 7 hours 20 minutes, time to acknowledge 12 minutes, SLA overrun 3 hours 17 minutes, above an investigation summary listing 14 findings in the signal
Case metrics
Time to acknowledge
12m
SLA overrun
3h 17m
Correlated signals
3
0
Attack chains correlated per device
0
Known attack patterns behind Attack Stories
0
Response verbs inside the case
0
Days re-read when an attack is confirmed
Why a case, not an alert

Alerts are not the hard part. Deciding is.

An investigation is deliberately separate from a risk, because it answers a different question. A risk asks what to fix. An investigation asks whether the behaviour is real at all, before anyone spends a morning on it.

One case per device, not one per detection

While a device has an open case, every new detection on it folds into that case, however long it has been open, and is listed under Detections in this case. Each detection there gets its own verdict (malicious, authorised or false positive), and one that is about something else can be moved to a case of its own. The headline and severity follow the most severe detection still in question. When an analyst closes a case as a false positive, benign or authorised, the same detection with the same evidence stays quiet for fourteen days. New evidence, a higher severity or a critical signal still opens a case. An urgent case names what opened it, such as "Mass file modification by Discord.exe", with the program, its services and the folders it wrote to.

Durations, not timestamps

Case age, time to acknowledge, time to contain and SLA overrun are shown as durations, which is the form an SLA is written in and an MSP reports on. Time to acknowledge measures the first human touch, not the detection that opened the case.

The next responder can tell what was done

Playbook steps carry completion state and a note recording what was actually carried out, rather than a read-only list. Dismissal reasons are presets, because a free-text box on its own produces "fp", "test" and blanks, which is worthless as tuning input. A step can carry a response action, such as isolating the host or signing the user out: ticking it runs that action through the same checks as the response panel, as the person who ticked it, an L2 investigator or above and never an L1. Only reversible actions that need no further input can sit on a step, and a failed action leaves the step open. A case without a playbook gets the best match when it is confirmed or escalated, and the trail records which one and why. Editing a playbook keeps the ticks and notes already made on it.

Correlation

A chain reads as a chain, not fourteen alerts.

Forty-one attack chains are checked on each device inside a 30-minute window. Twenty-two of them pair a real-time detection with a configuration change, so a cleared security log followed by antivirus switched off is one chain, not two unrelated cases.

  • Four chains open as Possible rather than Suspected, so the doubt is recorded on the case, not in a comment.
  • Findings and changes merge into one sequence, tagged with the technique the agent actually reported.
  • The case window is fixed, never sliding. Low and slow cannot hold it open.
  • Routine changes never build a chain. An updater, a rollout seen on three or more devices or a protection switched back on is marked Routine with its reason, and chains and stories skip it. The same change on three devices across the MSP's clients within a day also counts as a rollout, but suspicious persistence and critical chains never do.
Evidence collected by the agent's urgent uplink: 14 findings recorded across 4 finding types and 3 correlated signals, with a note that the batches fall between 08:25:45 and 08:41:12 inside a 30-minute correlation window, above an attack chain running from a registry persistence key through two file drops to a PowerShell process

Three correlated batches, one window, one case.

Attack Stories

Four alerts are one attack. The case should say so.

On every heartbeat, per device, Secuno groups what the agent reported by what the signals share (process lineage, file hashes, threat domains and addresses, files in user-writable folders) over up to 24 hours, configuration changes and behavioural deviations included, so an afternoon-long attack is one story rather than several. It matches the cluster against eighteen known attack patterns. The result is one story with a verdict (blocked, active or contained), a confidence, a plain-English summary, the MITRE chain and a timeline with one line per step. For a matched pattern it adds the sentence that matters most: what the technician should do next, not only what happened. The card shows the probable start, the step to watch for next, what joined later and why, and every part of the confidence, and lets you block a domain, sign a person out of Microsoft 365 or look for the hash on other devices from the same place. A story needs a real detection: routine administration, such as an administrator logon or a mapped drive, can add context but cannot start one.

The investigation is the case of record

A story at 70% confidence or with a strong score opens an investigation, and the story lives on it. The analyst's decisions on the case set the story's verdict, and an automatic closure never counts as a false-positive call. Three false-positive marks on one pattern quiet it for that client for ninety days, and never for any other client.

One attack across devices, one case

The same file hash or C2 domain on a second device makes one story, and lateral movement joins the story on the device it came from. When lateral movement, DCSync or roasting on a second device names the same user, the two stories merge into one cross-device incident, always critical, with the timeline tagged per device. The attack's next hop is not a second ticket.

A baseline that knows what normal looks like

Each device learns its process chains, the domains it resolves and its outbound connections from ordinary activity and from cases closed as a false positive or authorised, not only from alerts, and never records the address, the URL or the user. A deviation inside a story's window joins that story as evidence. Comparisons arm after fourteen days, and a new device can borrow a cohort of mature peers from the same client until its own baseline is ready: off, report only or raising cases, per client.

Dwell time

Detected today. Started when?

When an analyst confirms a story active or contained, Secuno re-reads that device's last ninety days knowing which signal types and tactics the confirmed attack used. Events that sat below the alert threshold at the time are scored again in that light.

  • A probable initial access date with a stated confidence of high, medium or low, the dwell period, and a retrospective timeline from probable entry to detection
  • A plain-English scope statement built from what the device could reach: sessions, shares, listening services and local administrators
  • The NIS2 Article 23 clocks, the 24-hour early warning and the 72-hour notification, counted from the probable first access alongside the same clocks from detection
  • When nothing in the window re-scores above threshold, it says so rather than guessing a date
Measured triage

Every open case says which way it leans.

Every close records one outcome: malicious, real but authorised, false positive or duplicate. Each night every detection gets a track record from those outcomes, per client and across the fleet, judged on a conservative bound. Every open case then carries a verdict: likely malicious, likely benign, or needs you.

Some things are never called benign

Credential theft, canaries, log clearing, tampering and indicator matches never get a likely-benign verdict, whatever the track record says. Only decisions by staff and partner roles count, and one device repeating the same benign event cannot carry a client's record on its own. A signed file is known by its path pattern and signer rather than its hash, so a dismissed auto-updater does not come back with every new version. Unsigned files keep their hash.

Auto-close is off until it is earned

Closing likely-benign cases automatically is a per-client switch, off by default. It applies only when every detection on the case is at least 95% benign at worst, over 50 or more decided cases. Any new signal sends a case back for triage.

Response times, measured from the activity

A Response times card shows median and 90th-percentile time to detect, to a first verdict and to containment, with Microsoft 365 and endpoint cases shown apart. Detected means when the activity happened, not when Secuno heard about it. Below five cases a figure is withheld rather than shown.

Response

Contain it without leaving the case.

Containment used to live only on the device page, behind a raw command dropdown. An analyst watching files being encrypted had to leave the investigation to stop it. The case now offers ten verbs, seven on its device and three on its Microsoft 365 account, and every one is mirrored into its activity trail.

  • Isolate or release the host, kill a process or its whole tree, quarantine a file, stop or disable a service, on this investigation's device only
  • Revoke sessions, disable or re-enable the account the case is about, on Microsoft 365 tenants that opted in. A case that groups alerts about more than one account refuses, and says why
  • Lateral quarantine as a lighter step than isolation, on Windows and Linux agents, from the device, the case or the attack story: internal traffic blocked in both directions, the internet kept, and a twelve-hour limit renewed by keep-alives
  • Learned workstation segmentation, opt-in per client: four weeks of learning who connects into each workstation, a proposal you review, then inbound enforcement ring by ring that pauses itself on a bad verdict, rolls back in one click and has a local kill switch
  • Every action is recorded in the case activity trail alongside analyst notes, status changes, assignments and escalations
  • Confirmed investigations escalate into a formal incident or a risk, carrying affected devices, responders, linked risks and tickets. The machine opens an incident on its own only for activity across devices; one device's detections stay one case. An incident runs open, contained, eradicated, recovered, closed, with a post-incident report. It has an owner, and its own playbook checklist that takes over the case's steps with their ticks. A deleted incident can be restored for thirty days
app.secuno.nl/investigations/INV-84
The response actions panel inside a Secuno investigation: isolate host, terminate process, quarantine file and collect forensics, noted as dispatched to the affected workstation and recorded in this investigation's activity
Dispatched to the device in the case, recorded in the case. No separate console, no raw command dropdown.
Guardrails

What the agent will do, and what it refuses.

The backend validates every queued command against an allow-list rather than accepting any string and leaving the agent to reject it. Eighteen command types can be queued at all, among them removing one exact software name and version (on Linux simulated first, and never an essential package, a kernel, the SSH server or the agent) and applying or reverting a vendor mitigation. Of those, these seven are the ones a case can send. The refusals are the interesting part: each one exists because the alternative is a support call at three in the morning.

Device commands a case can send
kill_processTerminates the process the detection saw, matched on its image and start time, so a recycled process number can never aim the kill at something elseRefuses PID 0, PID 4, the agent's own process, and any process whose image or start time does not match
kill_process_treeTerminates a process and every descendant, because the encryptor is often the child of what was detected and killing the parent alone leaves it runningRefuses any process, at any depth, it cannot confirm by identity and start time, and stops at forty processes
isolate_networkBlocks all inbound and outbound traffic except loopback, the agent itself, DNS and DHCP, installed in a single filtering transactionRefused outright if the Windows filter engine is unavailable or the agent cannot resolve its own executable path, and the console says why
unisolate_networkRemoves the isolation filters and clears the isolation recordNothing to refuse: this is the way back
quarantine_fileMoves the file into quarantine with a manifest recorded alongside itNever quarantines a file from the Windows system, WOW64 or component-store directories
stop_serviceStops a named serviceNever force-stops dependents, and never touches a protected service: the agent itself, boot and session integrity, the patch pipeline, Defender, the firewall, or DHCP and DNS
disable_serviceSets a named service's start type to disabledThe same protected-service list, checked before the order is queued and again on the device

Isolation carries a local time-to-live, renewed by every successful heartbeat. A device you can still reach stays isolated for as long as you need. A device you cannot reach rejoins the network rather than sitting cut off with no remote way back. Every heartbeat re-checks that each expected filter is still present in live kernel policy, and reports a repair as a tamper signal.

Detection content

New detections without a new agent.

Sigma and YARA rules reach devices as packs signed with a dedicated key, so a new detection does not wait for an agent release. Each rule runs live, in shadow or off, per client. A match is a detection like any other: it folds into the device's case, and Why this fired names the rule, its author and its references.

  • Sigma runs over every event-log channel the agent reads, plus Sysmon where it is installed, and a Sysmon that disappears raises a tamper case. On Linux it runs over every program start
  • YARA uses a built-in engine that agreed with reference YARA on every rule across 21,817 test files and 2,696 Linux binaries. It scans file drops, within a budget, and process memory
  • Public packs start conservative: the stable Sigma pack (32 rules) is on, while the test pack (2,218 rules) and the public YARA pack (2,919 rules) are off by default
  • A daily job promotes a quiet shadow rule after 14 days and demotes a noisy live one. A person's choice always stands
  • Community and client rules report at most high. A client's own Sigma rules are validated on save and never trigger automatic action, and a YARA file match never kills a process
  • A response rule library, switched on per client, covers attacks the built-in rules miss: PowerShell that destroys backups, dumps credentials or disables AMSI, and one live stable Sigma process-creation rule turned into a terminate rule. Every template installs report-only and can be armed only after fourteen days of observation, with every firing reviewed and none a false positive
Running a SOC service

Tooling for the SOC you sell.

For an MSSP that sells monitoring as a service, Managed SOC turns the contract into something the platform measures. A platform admin switches it on per client, and it runs on the same cases, verdicts and response actions described above. Two analyst roles, assigned per client, split the work for your own team: an L1 triage analyst works cases and raises tickets but never acts on an endpoint, and an L2 investigator adds response actions, isolation and incidents. An L1 escalation is a request: the case keeps its status, the client's L2 analysts are notified, and nothing opens until an L2 or above signs it off. The L2 can decline it and the L1 can withdraw it.

The contract, per client

Coverage is business hours in the client's own time zone, or 24/7. The contract sets which severities page, which of those still page out of hours, and targets to acknowledge, contain and resolve for each severity. Outside a 24/7 contract, the clock counts only covered hours unless the severity pages out of hours.

Rota and pager

Shifts are primary or secondary. The pager reaches the primary at once, the secondary at half the acknowledge target, then both plus the client's leads at the target: in-app, by email, or through PagerDuty, Twilio, Slack, Teams or a webhook. Gaps in the rota and people without a phone number are shown.

Figures against target

Detect, acknowledge, contain and resolve times are shown against target for the portfolio and per client, with every miss listed. A confirmed case stays in the SLA check and the priority queue, with a tighter deadline. Acknowledged means the first human action on the case. SOC-first routing can hold new alerts back from the client's own integrations, and logs every held event with the reason.

Every client can also set its own escalation policy: up to four steps for an untouched critical case or active story, to people or to Slack or Teams. Picking up the case stops the remaining steps.

Questions

The things worth asking first.

How is an investigation different from a risk?

A risk asks what to fix and carries an owner, a deadline and remediation steps. An investigation asks whether a behavioural signal is real at all. It runs new, triaging, then confirmed or dismissed, and a confirmed case escalates into a formal incident or a risk when remediation is actually required.

What happens when the same signal fires again after a case is closed?

That depends on how it was closed. When an analyst closed it as a false positive, benign or authorised, the same detection with the same evidence is counted on the closed case for fourteen days instead of opening a new one, so a known-good event does not come back every morning. New evidence, a higher severity or a critical signal still opens a case, and closes by auto-triage or an upstream system are not held. For something you know will recur, the dismiss dialog can also hold it for this device or for all of the client's devices, for up to ninety days and with a required reason. A hold covers only that detection family and the exact elements of the dismissed case, and never protected evidence: a canary, LSASS or DCSync access, shadow-copy deletion, a known-bad hash, an indicator match, an unsigned file or a new signer. Holds are listed, revocable and audited. Otherwise a new case opens rather than silently reviving the closed record, so the timeline never misreports when the activity began.

Can a slow attacker hold the correlation window open?

No. The window is a fixed length and deliberately does not slide. A sliding window can be held open indefinitely by a low-and-slow signal, and that is precisely the case where an accurate start time matters most.

If we isolate a device, can we lose it?

The isolation filters are installed in a single transaction that always carries the agent's own exception, so the host is never blocked without a way for the agent to talk back, and the command is refused outright if the agent cannot resolve its own path. Isolation then carries a local time-to-live renewed by each successful heartbeat, so an unreachable device rejoins the network instead of being stranded.

Can an obfuscated command line slip past the agent?

Command lines are read the way the shell reads them. Carets, quotes and variable tricks are undone and encoded PowerShell is decoded before anything is classified, and every rule runs over the raw, normalised and decoded forms. The same classifier runs whether the agent sees a process through its kernel trace or through the Security log, so weaker telemetry does not mean weaker detection. A process that lies about its parent is reported as parent PID spoofing with the real creator named, and stopping the agent's trace session is itself a detection.

How much of the queue is noise?

A September review of every open case traced most of the noise to detection defects rather than judgement calls: Windows' own database maintenance read as credential theft, builds and installers read as ransomware, JavaScript engines read as implants. Six sources like these were closed at the detection itself, and a case with one genuine finding among the noise stays open. Routine changes now open nothing: updaters, MDM enrolment, OneDrive, signed installers, Edge updates, the same signed file on ten or more of a client's devices, and routine NTLM, while a file burst that recurs on the same device is graded down a step. And a device with an open case gets one case, not one per detection.

Does it see an attacker moving between machines?

Yes, on Windows. An executable or script written into an admin share such as ADMIN$ or C$ opens a HIGH case. Remote service execution in the PsExec style, a network logon, a file dropped into the share and a service installed from it within two minutes, opens one confirmed HIGH case naming the address, account, file and service. A person's account logging on to a workstation from an address it has never used there is MEDIUM after fourteen days of learning, and HIGH when the account is a local admin or the source is another of the client's workstations. These need Windows file-share auditing switched on, which is often not the default, and the heartbeat reports whether it is.

What does an auditor get afterwards?

The case activity trail records every status change, assignment, escalation, analyst note and dispatched response action. Incidents add an owner, affected devices and users, responder assignment, linked risks and tickets, and a post-incident report with a PDF export, and every create, edit, status change, owner change, delete and restore is in the audit log. Investigations measure detection latency and time to first verdict over ninety days. Case reports print without cards splitting across a page boundary mid-evidence.

See it on your own fleet

Bring us a noisy endpoint.

The fastest way to judge a detection story is to watch it triage something real. We will walk a live case end to end, from the signals that opened it to the action that closed it.