Legitimate, Misconfigured, or Malicious? Classifying Bandwidth Hogs
The top-talker workflow ends with a name: this machine, this conversation, this much traffic. A name is not a verdict. The same flat-topped graph can be drawn by a backup job doing exactly what it should. It can also come from that job doing the right thing at the wrong hour. Or a compromised workstation might be quietly shipping your file server to a stranger. Those are three different problems with three different responses. Pick the wrong one and you either accuse an innocent server of crimes or politely reschedule a data breach.
The traffic carries its own evidence. Four signals separate the three verdicts: where the traffic is going, when it is happening, which direction it flows, and what shape it makes over time. None of them needs security training to read. They need a graph, a schedule, and the habit of checking all four before you decide. I have classified more hogs than I can count. And the ones I got wrong were the ones where I read one signal and stopped.
This guide teaches the four signals, walks through the classic culprits in each bucket, and gives you a table to classify against in a minute or two. It also says plainly where the signals stop. Traffic records can make a hog look suspicious; they cannot prove guilt. Knowing where classification ends and investigation begins is part of the skill.
In this guide
- Three verdicts, three responses
- The four signals
- The classic legitimate hogs
- The classic misconfigurations
- The malicious patterns — and an honest caveat
- The classification table
- Three quick walkthroughs
- When you still can't decide
Three verdicts, three responses
Every bandwidth hog you will ever catch lands in one of three buckets. Legitimate: the traffic serves a real business purpose and is behaving as designed. Think of the nightly backup, the all-hands video stream, or the new PC pulling its updates. The response is accommodation: schedule it, throttle it, or buy capacity for it, but do not treat it as an offense. Misconfigured: the software is legitimate but the behavior is wrong: wrong time, wrong scope, wrong rate. Usually, the cause is a default nobody reviewed or a setting nobody revisited. The response is a fix. These are the most satisfying tickets in the pile because the fix is usually one checkbox. Malicious: the traffic serves someone else's purpose — data theft, resource theft, or a machine conscripted into someone's botnet. A botnet is a crowd of compromised machines taking orders from one attacker. The response is a security process, not a network tweak.
Experienced admins will tell you the boring truth: the overwhelming majority of hogs are in the first two buckets. Backups, sync clients, and update storms outnumber attackers by a wide margin in ordinary networks. That is exactly why the third bucket is dangerous — after the twentieth backup job, it is tempting to stop looking carefully. I have written “probably the backup” on a ticket before opening the graph. It was the backup. That is not the point. The signals below exist so you do not have to choose between paranoia and complacency. You just read the evidence the same way every time. The verdicts can also nest: a legitimate job can be misconfigured, and a malicious process can hide inside legitimate software. Classify the behavior in front of you, not the reputation of the program producing it.
The four signals
Signal 1: Destination — where is it going?
The far end of the conversation is your strongest single clue. Some destinations vouch for themselves. These include your own backup server, your own replication partner site, and the address ranges of cloud services your company actually subscribes to. They also include the content-delivery networks that operating system and antivirus vendors use for updates. These are large distributed download services. A reverse DNS lookup asks DNS for the name behind an address. That or a quick address ownership lookup often tells you the mystery endpoint is a household-name cloud platform. It can take just seconds. At the other end of the scale sits the destination nobody can explain. The address has no meaningful name and belongs to a hosting provider you have never done business with. It is in a part of the world your company has no dealings with. Unknown does not mean guilty — but known-and-expected goes a long way toward innocent.
One wrinkle keeps this signal honest: attackers know that reputable clouds make good camouflage. Plenty of malware stages stolen data through the same big-name storage platforms your business uses. That is precisely so the destination looks boring. So read this signal narrowly. “A cloud service” vouches for little. But “the specific service we subscribe to, from a machine that has a reason to use it, doing the kind of thing that service does” vouches for a lot. The gap between those two statements is where the other three signals earn their keep.
Signal 2: Timing — when is it happening?
Traffic has office hours too. A large transfer during the business day competes with users and gets noticed. The same transfer at three in the morning is often exactly where it belongs. That is where sensible admins schedule heavy jobs. So the question is never “is 3 AM suspicious?” — it is “does this 3 AM traffic belong to a job I can name?” Your backup schedule, update schedule, and change log are the answer key. Off-hours traffic that matches a scheduled job is reassuring. Off-hours traffic that matches nothing on the calendar deserves a closer look precisely because someone may have chosen that hour hoping nobody was watching. And a scheduled job showing up at high noon is the signature move of the misconfigured bucket.
Signal 3: Direction — download or upload?
Big downloads are the background noise of a working network: updates, media, installers, people fetching things from the internet, which is what the internet is for. They cause congestion, but they rarely raise security questions. Big uploads earn more scrutiny, for a simple reason: an upload is your data leaving. Plenty of uploads are innocent — offsite backups, cloud sync, sending a finished project to a client, video calls pushing your camera feed out. But exfiltration is the security word for data being stolen off your network. By definition, it is an upload. So a large outbound flow to a destination you cannot vouch for combines the two worst answers on the sheet. Direction alone convicts nobody; it decides how carefully you read the other three signals.
Direction also changes how much damage a hog does. Many business internet connections are asymmetric: the upload lane is a fraction of the download lane. So an upload hog saturates with far less traffic than a download hog of the same appetite. It hurts everyone sooner. A modest-looking outbound flow can be the whole story of a miserable afternoon.
Signal 4: Duration and pattern — what shape does it make?
Zoom the graph out and read the traffic's silhouette. A scheduled burst is a mountain that rises at the same time on the same days and ends when the job ends. It is the shape of automation, almost always legitimate or misconfigured. A one-off mountain is the shape of a human doing something unusual once: a huge download, a first-time cloud sync. A constant drip is modest but unbroken traffic around the clock, weekends included. That shape deserves attention, because ordinary user activity rests when the users do. Steady around-the-clock flows belong to servers doing server things, or to software that does not want to be noticed. Think of a miner's continuous check-ins or a slow-and-quiet exfiltration deliberately kept under the alarm threshold. Or a sync loop might be endlessly re-sending the same files. The shape does not name the culprit, but it tells you which suspect list to open.
The classic legitimate hogs
Learn this cast of characters and you will recognize most hogs on sight. Backups are the heavyweight champions: they move everything you own, on a schedule, to a known destination, and they are supposed to be huge. Cloud sync clients are backups' chaotic cousins, well-behaved day to day. Then someone drops a folder of raw video into a synced directory. The client faithfully attempts to upload the whole thing at once. First-time syncs of a new account are the same story at larger scale. Update fan-out happens when every machine in the building independently downloads the same operating system or antivirus update from the internet. Dozens of copies of the identical bytes go through one pipe. The problem has a well-known fix: a local update server or peer caching. Then the bytes cross your internet link once. Video events include the all-hands stream and a training webinar. They put a copy of the same stream on every attendee's screen unless your platform multicasts or caches. Multicasting sends one copy that many screens share. And every network has its by-design bulk movers: imaging a lab of PCs, media teams moving footage, engineering syncing CAD files, database replication between sites. Legitimate does not mean harmless. A legitimate hog can still flatten your link. It means the fix is scheduling and throttling, not removal. That is the territory of the conversation after.
The classic misconfigurations
The second bucket is legitimate software with a wrong setting, and a handful of scenarios cover most of it:
- The backup that runs at noon. The schedule was set for overnight, then something moved it. It could be a failed run that resumed hours later, or a job queued behind another job. Maybe nobody noticed a timezone or clock change. Or a “temporary” manual run became the new schedule. The traffic is perfect; the clock is wrong.
- The sync loop. Two sync tools pointed at overlapping folders, or one tool fighting itself over a conflicted file, re-uploading and re-downloading the same data forever. The signature is volume wildly out of proportion to the actual data — the same files, endlessly.
- Unthrottled replication. Replication and backup tools almost all have bandwidth limits available — and almost all default to unlimited. The job is right, the destination is right, the rate flattens everyone else. One checkbox.
- Updates from the wrong place. Machines configured to fetch updates from the internet instead of your internal update server — or branch-office machines pulling across the WAN link instead of locally. Often the config was right until a subnet changed.
- The retry storm. A job that fails, retries immediately, fails again, forever — moving the same first chunk of data all night and succeeding at nothing. The giveaway is huge traffic with no completed result.
- The forgotten test. A trial of some tool, a one-time migration job, a demo VM — still running, still moving data, owned by nobody.
What unites the bucket: the traffic pattern looks legitimate except for one wrong dimension — wrong hour, wrong volume, wrong source, wrong repetition. When one signal is off and three are fine, think misconfiguration before malice. The forgotten test is the one to go looking for on purpose, because nobody will report it. Networks accumulate those the way couches accumulate coins.
Tip: when you fix a misconfiguration, write down the default that caused it. “Replication tool ships unthrottled” is not one machine's problem. The same default is probably live on every other install of that tool you own, waiting for its own incident. One noon backup, properly noted, can inoculate the whole fleet.
The malicious patterns — and an honest caveat
The third bucket is the smallest and the most important not to miss. Exfiltration reads as sustained upload to an unfamiliar destination, often at off hours, sometimes in one bold mountain, sometimes as a deliberate slow drip staying under every threshold. Cryptominers are malware that steals your electricity and CPU to mine cryptocurrency. They are modest on bandwidth but distinctive in shape. Expect constant small connections to the same few unfamiliar endpoints, around the clock, from a machine whose fans have opinions. Botnet membership means a compromised machine taking orders from someone else. It shows up as your machine doing things it has no business doing. Examples include bursts of outbound mail, scanning sweeps across addresses, and spikes of traffic to endpoints that rotate. The common thread is traffic that serves no purpose you can attach to the business, from a machine whose owner did not ask for it.
The caveat deserves its own paragraph: flow data raises suspicion; it does not convict. Traffic records show who talked to whom and how much — not what was said. An encrypted upload to an unfamiliar cloud host could be exfiltration or could be a perfectly legitimate tool your colleague adopted without telling IT. Classification tells you when to escalate, not whom to punish. The security-side use of flow evidence is covered properly in flow data for security. That guide explains what the evidence can support, what it cannot, and how it fits an investigation.
Tip: the moment you genuinely suspect malice, your job changes from network troubleshooting to evidence handling. Do not clean the machine, do not reboot it, do not block the traffic and move on — and do not announce your suspicion widely. Note what you observed and when, save the flow records, and hand it to whoever owns security incidents. In a small shop, that may still be you. But you are wearing a different, more careful hat, following your incident process.
The classification table
Here are the four signals laid against the three verdicts. No single row decides anything; read all four and see where the weight lands.
| Signal | Leans legitimate | Leans misconfigured | Leans malicious |
|---|---|---|---|
| Destination | Your own servers; cloud services you subscribe to; known vendor update networks | Right kind of destination, wrong instance (internet instead of local server, wrong site) | Unfamiliar host, anonymous hosting, no explainable business relationship |
| Timing | Matches a job on your schedule, or normal business-hours activity | A known job at an unscheduled hour (backup at noon) | Off-hours activity matching nothing on any calendar |
| Direction | Downloads; uploads to destinations you can vouch for | Either — just too much, too fast, or too often | Sustained uploads to destinations nobody can vouch for |
| Pattern | Scheduled bursts; one-off human-shaped events | Endless repetition, retry loops, volume out of proportion to the data | Constant around-the-clock drip; slow steady upload; rotating endpoints |
| Your records | Explained by a ticket, schedule, or known project | Job exists but settings drifted from what's documented | No owner, no ticket, no explanation from the machine's user |
The last row is the one beginners forget: your own documentation is a classification signal. A change log and a backup schedule turn “unknown 2 AM transfer” into “that's the offsite job” in ten seconds. That is one of many quiet payoffs of writing things down. The row cuts the other way too: a device nobody wrote down reads as malicious on every signal, whatever it actually is.
At Harbourview Clinic a steady overnight upload to a hosting range nobody recognized ran for a month, weekends included. It came from an address with no owner in the inventory. The duty admin blocked the destination at the firewall and closed the ticket, which is the step this guide asks you not to take. The upload did not stop; it turned into a retry storm. The device behind it was a security camera that the facilities manager had bought on a departmental card. Its default was to stream the car park to the vendor's cloud until it succeeded. IT learned the camera existed when facilities raised a ticket saying it had stopped working. It now records to a box in the comms cupboard, and the car park was quiet throughout.
Three quick walkthroughs
Case 1: the 2 AM mountain. Large outbound transfer, every night, same start time, destination is the address of your offsite backup provider, ends before anyone arrives. Destination: known. Timing: matches the backup schedule. Direction: upload, but to a vouched-for endpoint. Pattern: scheduled burst. Verdict: legitimate — log it as a known character and move on. (If it ever stops appearing, that is a different alert worth having.)
Case 2: the noon mountain. Same destination, same volume, same job — at midday, flattening the office. Destination known, direction fine, pattern fine, timing wrong. One signal off, three fine: misconfiguration. Sure enough, last night's run failed on a locked file and the job rescheduled itself to retry “as soon as possible,” which meant business hours. Fix the schedule's retry window, and the incident becomes a one-line note.
Case 3: the quiet drip. A workstation uploads steadily, not enough to congest anything, but continuously, including weekends. It sends to a hosting-provider address with no name anyone recognizes. It only surfaced because you were reviewing top talkers by total volume over the week rather than by momentary rate. Destination: unknown. Timing: always, which matches no human schedule. Direction: upload. Pattern: constant drip. Verdict: suspicion on all four signals — stop troubleshooting, start your security process, and let the investigation (not the flow report) decide what it actually is. It might still be an employee's personal cloud drive set to back up their machine. That is the point of the caveat: the signals tell you to look, not what you will find.
When you still can't decide
Plenty of hogs land between buckets. An upload to an unfamiliar destination might turn out to be a departmental tool nobody registered. An odd-hours transfer might have no ticket because the ticket was never written. When the signals genuinely disagree, use this order of operations. First, ask the owner, curiously — most mysteries dissolve in one question, and how you ask matters. Second, look at the machine itself. Identifying the program generating the traffic settles what the network alone cannot. The host-side techniques in finding hogs without flow data work just as well when you do have flow data. Third, let the risk asymmetry break ties: a misclassified backup costs you an awkward conversation; a misclassified exfiltration costs rather more. When upload-plus-unknown-destination is on the table and answers are not arriving, escalate and be pleasantly wrong.
And whatever the verdict turns out to be, feed it back into your notes. The destination you verified today is a signal you can read instantly next month. Classification gets faster every time you do it honestly.
Where to go next
Classification is the hinge of the whole discipline. Upstream of it sits the drill-down that finds the hog. Downstream sit the human conversation and the fix and the prevention work that keeps the same hog from starring in next month's incident. The full set of guides lives at Top Talkers: Who's Eating the Bandwidth. Read all four signals, every time, in the same order, and you will neither accuse an innocent server nor reschedule a breach.
Put this into practice with SysPeep Monitor
Every signal in this guide assumes you noticed the hog in the first place. SysPeep Monitor handles the noticing: it graphs utilization on every interface it discovers and learns what your links normally carry. It alerts you when one pins at its ceiling or breaks pattern. So classification starts minutes into the incident, not days. The free trial runs on any Windows machine.
Download Free TrialLearn About SysPeep