Chatty Devices: Finding IoT That Talks Too Much
Every IoT device on your network is talking to someone. The thermostat murmurs to its vendor's cloud. The cameras stream to the recorder. The TV in the conference room checks for something every few minutes, and nobody has ever found out what. Most of this chatter is harmless, and some of it is necessary. A small slice of it is a device talking too much, at the wrong times, or to the wrong people. That slice is how you find out a gadget has been compromised, misconfigured, or was never quite what the box said it was.
"Too much" has no universal definition, and that is the hard part. A camera sends more traffic in an hour than a thermostat sends in a year, and both are behaving perfectly. You cannot judge IoT traffic against a single yardstick. You can only judge each device against its own kind's normal. That means the real skill here is not reading alerts. It is knowing what normal looks like before the abnormal day arrives.
This guide builds that skill in order. It covers what a per-class baseline is and what it records, and the four traffic anomalies actually worth your attention. It explains how to judge the eternal "is this phone-home traffic okay?" question. It gives you a calm quarantine workflow for the day the answer is no. None of it requires logging into a single device, which is convenient, because with this population you usually can't.
In this guide
- Normal is per-class, not global
- What a baseline records
- Example class profiles
- The four anomalies worth an alert
- The phone-home question
- A worked example: the talkative sign
- The quarantine workflow
- Profiles as living documentation
Normal is per-class, not global
Think of each device class as having a vocabulary — the set of things it has any legitimate reason to say on a network. A thermostat's vocabulary is tiny: report a temperature, fetch a schedule, maybe pull a firmware update (firmware being the software baked into the device itself). In traffic terms that is a trickle of small exchanges with one or two vendor services, and long silences. An IP camera's vocabulary is one word repeated forever: a fat, steady, continuous stream to its recorder. A badge controller speaks in short bursts to one server every time a door opens. A digital sign pulls content on a schedule and idles between pulls.
Every one of those patterns is healthy for that class, and alarming for a different one. A camera-sized flow is normal from a camera and a five-alarm anomaly from a thermostat. Chatter at three in the morning is routine for a backup-hour signage refresh. It is deeply odd from a badge reader in a building that locked up at six. This is why generic "high traffic" alerts are useless against IoT. Any threshold loose enough to tolerate the cameras is blind to a thermostat behaving like a camera. Any threshold tight enough to catch the thermostat pages you all day about the cameras.
The fix is to hold each device to its own class's standard. That requires knowing the standard — a baseline: a recorded description of what this kind of device normally does on your network. Baselines sound like homework, and they are, but they are short homework. Most IoT devices are creatures of extreme habit, and their entire behavioral repertoire fits on an index card.
What a baseline records
A useful traffic baseline for a device class answers four questions. Who does it talk to? The list of destinations — internal servers, vendor cloud services — that the device contacts in normal operation. For most IoT this list has fewer than a dozen entries and barely changes. How much? The rough volume per day and the shape: steady stream, periodic bursts, rare trickle. Order-of-magnitude is plenty; you are distinguishing kilobytes from gigabytes, not auditing bytes. When? The rhythm: constant, business-hours, scheduled overnight. How? Which protocols and ports — does it speak encrypted web traffic to its cloud, a streaming protocol to a recorder, a proprietary port to a local server?
Where does the data come from? The best source is flow data — the per-conversation summaries your router or switches can export. Those summaries record exactly the who, how much, and when above without capturing any content. If you have flow visibility on the IoT segment, an afternoon of looking at a device class's conversations gives you the whole card. Without flow data you can still get the volume-and-rhythm half from per-port traffic counters on the switch. You can get a surprisingly good destination list from your DNS server. DNS, the Domain Name System, is the service that turns names into addresses. Most devices look up their destinations by name before talking to them. So your own DNS logs are a free ledger of who your gadgets are calling.
Written down, a finished card is refreshingly small, smaller than the leaflet that came in the box:
CLASS PROFILE: thermostats (6 units, IoT segment) talks to: two vendor cloud services (update + telemetry; names on file) volume: under 5 MB per day per unit, small exchanges rhythm: brief check-in every few minutes, otherwise silent ports: outbound encrypted web only; accepts nothing inbound notes: traffic doubles briefly during vendor firmware pushes last review: after most recent firmware update (see change log)
Tip: baseline a class when it first arrives on the network, while the vendor's installer is still answering email. "What should this thing talk to, and how much?" is a routine question during deployment week and an act of archaeology two years later.
Example class profiles
Here is what a set of per-class cards might look like for a typical building. Your numbers will differ, and that is fine, because the point is the shape:
| Device class | Talks to | Volume & shape | Rhythm |
|---|---|---|---|
| Thermostat / sensor | One or two vendor cloud services | Kilobytes per day, tiny exchanges | Periodic check-ins, long silences |
| IP camera | The NVR; occasionally vendor cloud | Gigabytes per day, steady stream | Constant, around the clock |
| Badge controller | The access-control server | Small bursts per door event | Business hours, near-silent nights |
| Digital signage / TV box | Content service, update service | Bursts at content pulls, idle between | Scheduled pulls, often overnight |
| Smart plug / power gadget | One vendor cloud service | Kilobytes per day | Periodic heartbeat |
Notice how little is on each card, and how sharply the cards differ. That contrast is the whole detection engine. Once these cards exist, "a signage box streaming steadily all day" or "a smart plug moving gigabytes" stops being an opinion. It becomes a documented deviation. Note also that the destination column names services, not addresses. Vendor clouds move between addresses freely, but the set of named services a device uses is remarkably stable.
At Meridian Parts a volume alert fired on an address in the IoT segment that had no inventory entry and no card. It was sending a small packet to a vendor service every four seconds, around the clock. That added up to more traffic in a week than the thermostats moved in a month. Three people spent the better part of three days on it, including a walk of every floor with a port map. Then it was traced to the vending machine in the canteen, which was reporting its stock levels. Every four seconds. The vendor confirmed this was by design, the card took ten minutes to write, and the machine has not been mentioned since.
The four anomalies worth an alert
With baselines in hand, four deviations cover nearly everything worth catching. New destinations: a device starts talking to somewhere that has never been on its card. The destination might be a service in a part of the world its vendor does not operate in, or a residential-looking address. Most tellingly, the destinations might be other devices on your own network it has no business knowing. Volume jumps: a device's daily traffic leaps an order of magnitude with no firmware update or configuration change to explain it. A thermostat that moved a gigabyte yesterday has a story to tell. Odd hours: activity falls outside the class's rhythm. The badge controller might be having a busy conversation at three in the morning. The signage box might be active continuously instead of at its scheduled pulls. Scanning behavior: the device starts knocking on many addresses or many ports in sequence — the network equivalent of rattling every doorknob in the building. Legitimate IoT almost never scans; compromised IoT almost always does, because finding neighbors to infect is the first job malware gives a captured gadget.
All four are visible in flow data, and none require seeing inside any packet. Turning flow records into security signals — the queries, the tooling, the investigation techniques — is its own discipline, covered in using flow data for security analysis. This guide's contribution is the IoT-specific baseline that tells you which of those signals matter for which class. Rank them for response, too. Scanning is drop-everything (it means the device is probably not yours anymore). New-destination and odd-hours are same-day investigations. A volume jump alone is a look-this-week unless it pairs with one of the others.
Set expectations for the first weeks: there will be false alarms, and each one is tuition. The signage boxes will trip the odd-hours rule until you learn their content schedule; the whole camera class will spike the day the vendor pushes firmware. Tune per class, not per device. When a false alarm teaches you something, the lesson goes on the class card, and the alert rule inherits it. After a month of this the rules go quiet, and from then on a firing means something. That arc — noisy, tuned, trustworthy — is the normal life cycle of every behavioral alert. Abandoning it during the noisy phase is the only way to fail at it.
The phone-home question
Sooner or later every admin stares at a flow record and asks the question this whole subject orbits. The question is: this gadget is talking to some cloud service I do not recognize — is that fine, or is that an incident? Vendor telemetry (the device reporting its own health and settings back to its maker) is genuinely normal. Modern devices check for updates, report health, and sync settings with their maker's cloud. The skill is separating expected telemetry from traffic that deserves escalation, and three tests do most of the work.
Destination consistency. Legitimate telemetry goes to a small, stable set of vendor-operated services, usually with names that plausibly relate to the vendor. Every unit of the same model talks to the same set. Walk the flows for all six of your thermostats: if they all speak to the same two services, that is a vendor pattern. If one of the six also speaks to somewhere the other five have never contacted, that one device — not the vendor — is your question. Volume proportionality. Telemetry is small: settings, health pings, update checks. Kilobytes to low megabytes per day. When the volume looks like payload, the traffic has outgrown every legitimate purpose the device has. Examples are a sensor uploading gigabytes or a badge reader streaming steadily. For traffic that looks like payload, size alone makes it suspicious. Vendor documentation. Reputable vendors will tell you — sometimes in an install guide, sometimes only when asked — which services their device needs to reach and why. Ask, and file the answer with the class profile. A vendor may be unable or unwilling to name their own endpoints (the services the device needs to reach). In that case, they are telling you something about how much traffic scrutiny their product will survive. That answer belongs in your next purchasing conversation.
Run the three tests in that order and most mysteries resolve in minutes. If the destination is consistent, the volume proportional, and the traffic vendor-confirmed, write it on the card and move on. Fail two of the three, and you have earned the quarantine section below.
A worked example: the talkative sign
Monday morning, your monitoring flags that the lobby's digital sign moved far more traffic overnight than its card allows. You pull the flow summary for the device and lay it against the profile:
CLASS PROFILE says: content pulls overnight, bursts, ~200 MB, two known services FLOWS for sign-lobby-01, overnight: to content service (known) 180 MB scheduled pull <- normal to update service (known) 2 MB version check <- normal to unknown external service 3.1 GB steady upstream <- ??? to 40+ internal addresses, port sweep, small probes <- !!!
The annotations write themselves. The first two lines match the card — ignore them. The third fails two phone-home tests at once. The destination is on no card, and the volume is payload-sized and flowing the wrong direction. Signs download content; this one is uploading three gigabytes. The fourth line is the clincher: a port sweep across your internal network is scanning behavior, which content players never legitimately do. Cross-checking the other signage units shows their overnight flows match the card exactly, so this is one compromised device, not a vendor change. That combination — scanning plus unexplained upload — skips the deliberation. Isolate now. Verify the details after. That is precisely the exception the workflow below carves out. Total time from alert to switch-port quarantine: about ten minutes, most of it spent finding the sign in the inventory. The device never answered a single question — the network answered all of them.
The quarantine workflow
The day a device fails the tests, the goal is a calm, ordered response — not a panicked unplugging spree. Nor is it the opposite failure of watching a compromised gadget for a week while it explores your network. The tool for the second step is a quarantine VLAN: a virtual LAN. That means a separate network segment carved out on the same switches, with no route to anything. Four steps:
- 1. Verify. Confirm the anomaly from a second source before acting. The flow record says the sign is scanning — do the switch-port counters agree traffic surged? Is it one device or the whole class? One device misbehaving suggests compromise; a whole class changing together suggests a firmware update or vendor-side change. Identify the exact device, its location, and its owner from your inventory — which is why the inventory exists.
- 2. Isolate. Cut the device's network access without destroying evidence. Move its switch port to a quarantine VLAN with no route to anything, or block it at the firewall and switch. Prefer isolation over powering it off. A powered device can still be inspected and keeps its volatile state. Yanking power on something like a door controller has physical-world consequences. Do check what the device controls before isolating it, and loop in its owner if cutting it off stops badge doors or safety equipment.
- 3. Contact. Tell the device's owner what you did and why, and open a case with the vendor. Here is the traffic we observed. Here is when it started. Is this expected behavior, and what is the remediation? Sometimes the answer is embarrassing-but-fine ("that is our new update service"). Sometimes it is a known issue with a fix. Sometimes your report is how the vendor learns about a problem. All three outcomes beat guessing.
- 4. Document. Record the timeline — what was observed, when, what was done, what the vendor said, how it resolved. Update the class profile with anything learned. If the resolution was a firmware fix, note the behavior change. If the traffic turned out legitimate, add the destination to the card so the next admin does not re-fight this fight.
One prerequisite makes step 2 a five-minute action instead of an emergency project. The quarantine VLAN has to exist before the bad day. That is part of why IoT belongs on its own segment in the first place. The segmentation and visibility strategy for guest and IoT networks is covered in monitoring guest and IoT networks. And one exception to the calm ordering: a device actively attacking others — scanning hard, hammering a server — gets isolated first and verified second. Containment beats process while the attack is still running.
Profiles as living documentation
The class profiles are not a one-time project; they are living documents with a maintenance rhythm, and the rhythm is light. Review each card when something changes on the vendor side. A firmware update is the classic trigger, because updates legitimately change destinations and volumes. An un-updated card turns the next update into a false alarm. Review on your own trigger too. When a new class of device arrives, its card gets written during deployment. Fold the question into your standard intake for new systems. When an incident teaches you something, the card absorbs the lesson. Once a year, skim the whole stack against reality. During that review, retire cards for device classes that left the building, and spot-check one live device per class against its card.
Keep the cards where the next responder will actually find them — with your network documentation, linked from the device inventory entries. Keep them short. A card that fits on a screen gets read during an incident; a ten-page behavioral analysis gets skipped in favor of guessing. The card is also a quietly powerful purchasing tool. After a few of these, you will notice which vendors' devices have small, stable, documented vocabularies. You will notice which ones chatter unpredictably to mystery services. That knowledge, written down, is how the next generation of gadgets entering your building gets a little less chatty than the last.
Where to go next
The method travels well beyond thermostats. Baseline by class and alert on the four deviations. Judge phone-home traffic by consistency, proportion, and documentation. Quarantine calmly with the VLAN you built in advance. Do that and the devices stop being a crowd of strangers and become a short list of habits you can check in an afternoon. You may even find out what the conference-room TV has been checking for. The same outside-in philosophy — watching sealed devices from equipment you control — runs through the rest of IoT, Cameras & Building Systems. Start with the five observable signals for unmanageable devices, then apply it to building systems and vendor appliances.
Put this into practice with SysPeep Monitor
Baselines start with measurement, and measurement starts with your switches. SysPeep Monitor discovers the devices on your network and reads per-port traffic over SNMP. It lets you set per-device thresholds. So the thermostat that suddenly moves camera-sized traffic trips an alert instead of a hunch. Dashboards and reports make last month's normal easy to look up. Free trial on any Windows machine.
Download Free TrialLearn About SysPeep