Home › Topics › Flow Monitoring (NetFlow, sFlow, IPFIX) › Flows for Security

Flow Data as a Security Sensor


You probably set up flow monitoring to answer capacity questions. Who is filling the internet link? Did the backup overrun? What actually is all that traffic? Here is the pleasant plot twist: the same records double as one of the cheapest security sensors you can own. An intruder can encrypt every payload and disguise their tools as ordinary software. But there is one thing they cannot avoid doing — making network conversations. And flow data is the log of every conversation your network has.


Honesty before enthusiasm: flow data is a complement to real security tooling, not a replacement for it. It will not match malware signatures like an antivirus. It will not correlate events like a SIEM (security information and event management system — the platform that gathers and cross-references security logs). And it will not inspect traffic like an NDR (network detection and response) product. What flows offer is humbler and still genuinely valuable. They offer familiar reports in which certain attacker behaviors make recognizable shapes, plus a rewindable history of who talked to whom. For a small shop with no security operations center, that is a great deal more than nothing.


This guide shows you the six shapes worth knowing. It covers what each pattern looks like in ordinary flow views and why attackers produce it. Just as important, it covers the innocent explanations that produce the very same shape. Most alarms turn out to be a backup agent doing its job. By the end you will know what to glance for weekly and what flow data can never tell you. You will know exactly what to do on the day a pattern refuses to explain itself.


In this guide

  • Why flows make a good security sensor
  • The six patterns worth watching
  • The sampling caveat
  • Three starter moves for a small shop
  • What flows cannot do — and when to escalate
  • The pattern table

Why flows make a good security sensor


A quick recap: a flow record is a summary of one network conversation — source address, destination address, protocol and ports, start time, duration, bytes, and packets. It is the itemized phone bill of your network: every call logged, no recording of what was said. Detectives have closed cases on phone bills alone, because who called whom, when, how often, and for how long is astonishingly revealing. The same logic applies to network conversations.


Three properties make flows unusually well suited to the security job. First, the records are made by your network gear, not by the endpoint. Malware that hides its processes from the operating system it infected cannot reach into a router and un-make the conversations it had. Second, coverage is broad: one export point in the right place sees every host behind it, including the printer nobody patches. Third, and most underrated, flow data is historical. A warning may arrive from somewhere else — an odd email, an antivirus alert, a notice from your ISP. Flows then let you rewind and ask what that machine has been doing for the past two weeks.


The six patterns worth watching


Each pattern below is a shape: something you can recognize in ordinary top-talker, conversation, and time-series views without any special software. For each one you get the signature, the reason attackers produce it, the innocent look-alikes, and a sensible first check. One discipline keeps all of this honest: investigate before you accuse. A pattern is a reason to look closer, never a verdict by itself.


Pattern 1: Beaconing — the metronome

Malware that has settled into a machine needs instructions. So it phones its command-and-control server (often shortened to C2 — the attacker's coordination point) on a schedule. That might be every five minutes or every hour, sometimes with random jitter added to seem less mechanical. The signature is small, regular, patient: a few hundred bytes to a fixed destination, again and again, at metronome spacing, around the clock, for weeks.

Regularity is the tell, and the reason is human: people are bursty. Real browsing is a storm of flows, then silence, then another storm at no particular interval. Software checking in for orders is periodic in a way no human ever is. You will never notice a beacon in a volume view — it is far too small to rank. But filter a suspect host down to a single external destination and plot its flows across a day. A beacon then looks like a picket fence: identical marks at identical intervals.


Start     Source      Destination    Port  Bytes
02:00:11  10.0.4.31   203.0.113.9    443   1,342
02:05:11  10.0.4.31   203.0.113.9    443   1,342
02:10:12  10.0.4.31   203.0.113.9    443   1,398
02:15:11  10.0.4.31   203.0.113.9    443   1,342

The innocent look-alikes are numerous, because legitimate software beacons constantly. Examples are antivirus checking for updates, operating-system telemetry, and chat apps keeping sessions alive. Others are monitoring agents reporting in and time synchronization. Legitimate beacons go to destinations you can identify — vendor networks, well-known services. They appear on many of your machines identically, because everyone runs the same antivirus. One lone host beaconing to an address nobody can name is the interesting case. First check: look up the destination, then see whether other hosts on your network talk to it too.


Pattern 2: Internal scanning — the fan-out

An intruder's first move inside a network is reconnaissance: try addresses, try ports, see what answers. In flow data this makes a fan-out. One source produces a burst of tiny flows to many distinct destinations or ports in a short span. There are dozens or hundreds of conversations, most carrying almost no bytes because nothing answered the knock. In a conversation view it looks like a star: one center, many spokes, all thin.

Note the inversion from everything bandwidth taught you. This pattern is invisible in byte terms, because the volumes are trivial. But it is glaring if your collector can rank sources by number of flows rather than by volume. If your reporting tool has a "most sessions" or "top flow creators" view, this pattern is what that view is for.

The innocent look-alikes matter, because scanning your own network is a legitimate industry. Your vulnerability scanner scans. Your monitoring system's auto-discovery sweep scans (discovery is, technically, a very polite scan). Asset-inventory agents probe, and even some printers and media software search the subnet for peers. First check: is the source a system whose job is scanning? A fan-out from the monitoring server is Tuesday; the same shape from the receptionist's PC deserves your full attention.


Pattern 3: Upside-down volume — the exfiltration shape

Workstations download. That is their nature: pages, updates, mail, video — the traffic flows overwhelmingly toward them. So a workstation that suddenly uploads gigabytes to an outside address has turned its normal shape upside down. Data theft (exfiltration, in security vocabulary) looks exactly like that. It is large, sustained, outbound, often to a destination the host has never used before, often at night. The same inversion applies to servers. An internal database server that has only ever received connections may suddenly initiate a large outbound transfer to the internet. That is upside down twice — wrong direction and wrong role.

The innocent explanations are real and frequent. They include a newly installed cloud-backup agent making its enormous first upload, and photo or file synchronization catching up. They include a user legitimately sending a huge file for work, and a virtual-machine image pushed to a cloud service. First check: check the destination and the hour. A recognized backup provider at 1 AM reads very differently from an unknown host at 1 AM. Then simply ask the user. Most mysteries end in that conversation, amicably.


Pattern 4: Lateral oddities — peers that should not chat

On a typical office network, workstations talk to servers and to the internet — almost never to each other. An attacker expanding from one compromised PC to its neighbors (the industry term is lateral movement) breaks that rule. Workstation-to-workstation conversations appear, on exactly the ports built for administration and file access. Those include Windows file sharing on 445, remote desktop on 3389, and remote-management ports. A quiet client subnet suddenly developing internal chatter is worth attention. It is urgent when the chatter chains: one host starts conversations with several peers, and then those peers begin starting their own.

There is one caveat before you go looking. Your export point has to see east-west traffic (traffic moving sideways between internal hosts) before it can record any. Flow export configured only at the internet edge sees nothing that passes between two workstations on the same switch. That gap is one of the classics described in The Eight Blind Spots. The fix is exporting from a core or distribution switch as well as the edge.

Innocent look-alikes include screen-sharing and collaboration tools that connect colleagues directly, and a desktop quietly sharing a printer or a folder. They also include peer-to-peer update features that spread patches machine-to-machine. First check: which ports, and whether the pattern spreads. One stable pair on a screen-sharing port is probably two people collaborating. A growing web of connections on admin ports is not.


Pattern 5: After-hours activity from quiet hosts

Most office machines keep office hours, and their flow history is a tidy nine-to-five heartbeat: busy all day, silent all night. Attackers prefer the other hours, when nobody is watching. A workstation that is normally silent overnight suddenly moving data at 3 AM is a weak signal by itself. But it is a strong multiplier for every pattern above. Beaconing at 3 AM, a fan-out at 3 AM, a large upload at 3 AM: each is considerably more interesting than the same shape at 3 PM.

Innocent look-alikes dominate this category, which is why timing alone should never trigger an accusation. Examples are patch windows, scheduled antivirus scans, and backup agents. Others are cloud sync catching up after a busy day, and users who genuinely log in remotely at odd hours. First check: what the host actually did — volumes, destinations, ports — not merely the fact that it was awake.


Pattern 6: New-destination novelty

Servers are creatures of habit. A mail server, a database server, an application server — each talks to a short, stable list of partners, month after month. Each has the loyalty of a regular at a diner. That stability makes novelty cheap to spot. A stable server suddenly holding conversations with an address it has never talked to before is a genuine event, even at tiny volume. Notice that this check only works because flow data is historical. "Never talked to before" is a claim about the past, and you have the past on disk.

Innocent look-alikes include a newly integrated cloud service and a vendor moving its endpoints to new addresses. Others include an administrator doing maintenance from the console, and a licensing or certificate check added by an update. First check: what changed — deployments, updates, new software, new integrations — in the window when the new destination first appeared. Ticket history usually resolves this one in minutes.


Tip: keep a short "explained" list. Every time you chase a pattern and find an innocent cause, write it down with a date. The innocent cause might be the discovery sweep, the backup agent's beacon, or the peer-to-peer update feature. Next week's review gets faster. Over time the list becomes a map of your network's legitimate oddities: the noise you have already understood. So only the new noise demands attention.


The sampling caveat


Everything above assumes your flow records describe every conversation. Your export may be sampled: the device examines only one packet in every N and estimates the rest. This is standard for sFlow and common on busy routers. With sampling, detection quality changes, and not uniformly. Volume anomalies survive sampling well. A multi-gigabyte upload consists of so many packets that plenty of them land in any sample. So the upside-down-workstation shape still shows clearly. Beaconing survives poorly at small scale: a beacon may be a handful of packets every five minutes. At one-in-a-thousand sampling most check-ins contribute no sampled packet at all. In that case, the picket fence simply never appears in your data. Scan fan-outs land in between: a large, noisy scan still shows its shape through sampling; a slow, patient one slips beneath it.


Two practical conclusions follow. On a small network, prefer unsampled export at the edge if your hardware allows it — the record volume is usually manageable there. And wherever sampling is in effect, treat the absence of small-flow patterns as "not visible," never as "not happening." The full accounting of what sampling preserves and destroys is in Sampling: What 1-in-1000 Does to the Truth.


Three starter moves for a small shop


You do not need a security team, a budget, or new software to get value from any of this. Three moves, in order of payoff.


Move 1: keep enough history to look backward. Small organizations rarely discover compromises themselves. The news arrives from outside — a provider's abuse notice, a customer's puzzled email, an antivirus alert on a single machine. The first investigative question is always the same: what has that machine been talking to, and since when? Flow data answers it beautifully — but only if you still hold records from before the alert. This is the flight-recorder role. Nobody watches a flight recorder live. Its entire value is that it was already running before anyone knew there would be questions. Retention measured in weeks rather than days is the difference between an investigation and a shrug. (Retention costs disk; the storage side of this pillar covers that math.)


Move 2: a weekly odd-patterns glance. Ten minutes, protected by a recurring calendar slot. Scan the week for the six shapes: any workstation among the top uploaders? Any host with a startling flow count? Any new names in the servers' destination lists? Any internal chatter between clients? Any activity from machines that should have been asleep? You are not trying to catch everything. You are building the familiarity that makes the genuinely odd item jump out, and feeding the "explained" list as you go.


Move 3: simple threshold alerts, where your collector supports them. The volume and direction patterns threshold well precisely because they are big and directional. Alert when a client-subnet host uploads more than some generous daily figure to the internet. Alert when any single source exceeds an unusual flows-per-minute burst. Alert when a host in a server subnet initiates outbound traffic on ports it never uses. Start loose and tighten slowly. A security alert that cries wolf daily will be ignored by Friday, and an ignored sensor is just decoration.


What flows cannot do — and when to escalate


The limits deserve their own section, because acting beyond them causes real harm. Flow data has no payloads. It can say 40 GB left the file server for an unfamiliar address at 2 AM. It cannot say whether that was theft or a new replication job. It renders no verdicts: there are no signatures, no malware names, no "malicious" flag — only shapes and your judgment. And it holds no endpoint truth: it cannot tell you which process or which user account created the traffic. That is the job of endpoint tools and their logs. A flow pattern is grounds for suspicion, never proof of guilt. And the limit cuts both ways: clean-looking flows do not prove a clean machine, either.


Warning: never confront a person with an accusation built on flow data alone. The records prove that a machine moved bytes — not which human directed it, which program generated it, or why. Frame every conversation as "help me understand what this machine is doing." Half the time the machine is doing it without its owner's knowledge anyway.


So what happens on the day a pattern survives every innocent explanation? What about the lone beacon to an address nobody recognizes, or the workstation that really did upload half the file server overnight? You stop being the analyst and start being the first responder, in this order:


  • 1. Contain first, carefully. If the pattern suggests active data theft, isolating the machine from the network beats continuing to watch it perform. Disconnect or quarantine it — but do not wipe it, reimage it, or start "cleaning." The machine itself is now evidence, and evidence you scrub is evidence gone.
  • 2. Preserve the flow data. Export the relevant records and reports to files immediately, while full detail still exists. Collectors age old records into summaries on a schedule that does not care about your incident; the detail you need has a shelf life.
  • 3. Write the timeline. First seen, destinations, ports, volumes, and any other hosts showing related patterns — one page, in time order. It is the first artifact any responder will ask for.
  • 4. Hand it to the security function. Your internal security team if one exists; your managed service provider or an incident-response firm if not. This is not defeat — it is the design. Flows found the smoke; specialists handle the fire. Arriving with a timeline and preserved exports turns their first day of work into their first hour.

The pattern table


The six shapes on one page — signature, look-alikes, and the first check that separates them.


PatternSignature in flow dataInnocent look-alikesFirst check
BeaconingSmall flows to one destination at metronome intervalsUpdate checks, telemetry, chat keep-alives, agentsIdentify the destination; do other hosts talk to it?
Internal scanOne source, many destinations or ports, near-zero bytesVulnerability scanner, monitoring auto-discoveryIs scanning this source's actual job?
Exfiltration shapeWorkstation or quiet server suddenly uploading heavilyNew backup agent, cloud sync, legitimate big sendDestination and hour; then ask the user
Lateral movementWorkstation-to-workstation flows on admin and file portsScreen sharing, peer update features, shared printersWhich ports, and does the chatter spread?
After-hours activityNormally quiet host active in the small hoursPatch windows, scans, backups, remote workersWhat it did — not just when it was awake
New destinationStable server talking to a never-before-seen addressNew cloud service, vendor address change, maintenanceWhat changed when the destination first appeared?

Where to go next


Flow data will not replace a security stack. What it gives a small shop is eyes it did not have. It offers six recognizable shapes, a rewindable history, and a calm procedure for the day something real appears. The rest of Flow Monitoring (NetFlow, sFlow, IPFIX) builds the foundations this guide leans on. Those include what flow data reveals in day-to-day operations. They also include setting up export at the points that see the traffic you care about. And they include what sampling does to the patterns you now know how to read.


Put this into practice with SysPeep Monitor

Traffic patterns are one layer of awareness; the ground truth underneath is device and service health. SysPeep Monitor keeps watch on the basics that security work leans on — device health, interface status, and utilization over SNMP and WMI. Its thresholds and alert rules are built to keep the noise down, so a real anomaly is not buried under false alarms. The free trial installs on a Windows machine in minutes.

Download Free TrialLearn About SysPeep