Outbound Baselines: Knowing Your Normal Egress
Every guide about spotting data theft eventually says the same thing: "look for traffic that is unusual for that host." That advice is correct and, on its own, useless — because it assumes you already know what usual is. Most admins do not. They know the internet link is "about half full most days" and that "the backup runs at night." That is roughly where the knowledge ends. Ask them how much the finance file server sends to the internet on a typical Wednesday and the honest answer is a shrug.
An outbound baseline is the cure for the shrug. It is a written-down description of how much data each host (or group of hosts) normally sends out of your network. It records to how many places, to which places, and at what hours. Once you have one, "unusual" stops being a feeling and becomes a comparison: today's number against the baseline's number. That comparison is the foundation under every other guide in this pillar — odd-hours rules, new-destination events, the investigation playbook. All of them assume you can say what normal looks like.
This guide shows you how to build one from data you almost certainly already have. You will learn which three numbers to record per host and what to expect from each kind of device. You will learn how to treat first-seen destinations as events and how to think about destination reputation without buying anything. You will learn how often to refresh the whole thing so it stays honest.
In this guide
- What an outbound baseline is (and is not)
- Two lenses: per host and per segment
- Role-based expectations
- Day, night, weekday, weekend
- The "new destination" concept
- Destination reputation as a concept
- Building the baseline from flow data
- A baseline table template per role
- How often to refresh baselines
What an outbound baseline is (and is not)
A baseline is not a single threshold. "Alert if any host sends more than five gigabytes" is a threshold. It is wrong for almost every host at once: far too high for a receptionist's PC, far too low for the backup server. A baseline is a small profile, recorded per host or per group, made of three numbers and one list:
- Bytes out per day — the typical daily total and the highest total you have seen in a normal week (the "high-water mark"). Record both; the gap between them tells you how bursty the host is.
- Distinct external destinations per day — how many different outside addresses the host talked to. Workstations wander; servers are loyal to a handful of places.
- Working window — the hours in which the host is normally active, and whether it has any legitimate night or weekend activity.
- Top destinations — the short list of external addresses or names that account for most of the host's outbound bytes. For most hosts this list is surprisingly short and surprisingly stable.
That is the whole artifact. It fits in a spreadsheet row. The general craft of baselining — why averages lie, why you want ranges rather than single values — is covered in the Baselining guides. This article applies that craft to one specific question, which is what leaves.
A baseline is also not a promise that everything inside it is safe. A slow leak that has been running since before you started measuring will be in your baseline, looking perfectly normal. That is an unavoidable limitation, and the refresh section at the end has a partial answer to it. For now, accept that a baseline detects change, and change is most of what you are looking for.
Two lenses: per host and per segment
You will not baseline every device individually, and you should not try. Use two lenses. The per-host lens is for devices that matter individually. That means servers, anything holding sensitive data, and anything with a special job (backups, mail relay, monitoring). It also means a handful of workstations belonging to people with unusual access — finance, HR, executives, administrators. Each of these gets its own row.
The per-segment lens is for the crowd. A segment here means a group of devices that share a role and usually a network. Examples are the office workstation VLAN, the guest wireless, the printers, and the cameras and building systems. For these you record the group's total outbound bytes per day. You also record — this is the useful part — the typical per-device figure within the group. Workstations on the same VLAN behave alike enough that the group's typical value is a fair baseline for any single member. A workstation sending ten times the group's typical daily total is worth a look even if nobody has ever baselined that particular PC.
The segment lens also catches something the host lens cannot: a whole group drifting. If the workstation VLAN's outbound total doubles over a month with no obvious business reason, no single host will trip an alarm, but the segment will.
Role-based expectations
Before you have measured anything, you already know roughly what each kind of device should do, because the job dictates the traffic. Writing those expectations down first gives you something to compare your measurements against, and it makes surprises during the first measurement week obvious. Here is a starting set. Every value is relative, because absolute numbers depend entirely on your organization's size and habits.
| Role | Bytes out | Direction | Destinations | Night / weekend |
|---|---|---|---|---|
| Office workstation | Low. Attachments, video calls, background sync. | Downloads far more than it uploads. | Many, but small; a few large (your cloud services). | Nearly silent apart from patching. |
| Power user / media / developer workstation | Medium, bursty. Large uploads happen. | Still download-heavy on average. | Medium; includes code hosts, media platforms. | Mostly silent; occasional late work. |
| File server | Low externally. Its big traffic is internal. | Serves inward; should rarely push out. | Very few: updates, perhaps an off-site sync. | Sync or replication windows only. |
| Backup target | High, on schedule. | Upload-heavy by design. | One or two: the off-site or cloud target. | Almost all of its traffic is at night. |
| Mail relay | Medium, steady, follows business hours. | Roughly balanced. | Many, changing daily (recipient servers). | Low but never zero. |
| Public web or application server | Medium to high, follows visitor patterns. | Upload-heavy (it serves pages). | Many, changing (visitors). | Follows the audience, not the office. |
| Database server | Very low externally. | Should almost never talk outward at all. | Updates only. | None. |
| Printer, camera, building system | Very low. | Tiny check-ins; cloud cameras stream out by design. | One or two vendor addresses. | Constant, small, unchanging. |
Two rows deserve emphasis because they are where the worst surprises live. A database server should have almost no outbound story; any meaningful egress from one is worth a conversation the same day. And a file server is easy to misjudge because it moves enormous volumes internally. So its interface graphs look busy — but its external outbound should be tiny. Baseline the external half separately, or you will never notice when it changes.
Day, night, weekday, weekend
A single daily total hides the pattern that matters most for exfiltration, which is when. Split every baseline into at least three windows: working hours (for that host's users or its scheduled job), off hours on working days, and weekends and holidays. For a workstation, the working-hours window carries nearly all of its outbound bytes. Off hours carry patch traffic and sync leftovers. And weekends carry almost nothing. For a backup target, the picture is inverted. Record the typical and high-water values for each window rather than for the whole day.
This is more bookkeeping, but it is what makes the cheapest exfiltration signal work. "This workstation sent four gigabytes today" is ambiguous. "This workstation sent four gigabytes between two and three in the morning, against an off-hours baseline of a few megabytes" is not. The guide on Odd-Hours Transfers turns that split into alert rules. This guide just makes sure you have recorded the numbers those rules need.
Watch for weekly rhythm too. Many organizations have a heavy day — a reporting deadline, a weekly export to a partner, a sync that runs on Fridays. Note the heavy day in the baseline so the person reading an alert on that day knows it is expected.
The "new destination" concept
Of all the ideas in this guide, this is the one to take away if you take only one. A first-seen destination is an external address (or name) that a given host has never contacted before, as far as your records go. Treat its first appearance as an event. That means something you log, with a timestamp and a byte count, the same way you would log a login or a configuration change. Not an alert, necessarily; an event. Events can be reviewed, counted, and combined with other signals later.
Why it works: legitimate traffic has a long memory. The cloud services your company uses were first seen long ago and are seen every day. Update servers, partner systems, the news sites people read at lunch — all old friends. Genuinely new destinations do appear (a user visits a new website, a vendor changes an address), but they arrive small. A destination that is first-seen and immediately receives a large upload is rare, and rare is what you want.
Build it with a simple table that you update from flow records or firewall logs on a schedule. Keep one row per host-and-destination pair, with the first time seen, the last time seen, and the running total of bytes out. When a pair appears that is not in the table, that is your event. Keep two flavours: new for this host (a common, mild event) and new for the whole network (rarer, stronger). To keep churn down, group destinations by the network that owns them rather than the exact address wherever you can. That way, a large service rotating between addresses does not generate a fresh "new destination" every morning.
Tip: the first week of new-destination logging will feel like a firehose, because everything is new to an empty table. That is expected. Let the table fill for two or three weeks before you read the events for meaning. Resist the temptation to alert on them until the daily count of new pairs has settled into a low, boring number.
Destination reputation as a concept
Once you are logging destinations, you will want to sort them into rough categories. That is because "new" is more or less alarming depending on where the traffic went. You do not need a commercial reputation feed to start. Three home-made buckets do most of the work:
- Known-good: the services your organization actually uses — your email and file-sharing provider, your update sources, your business partners, your own cloud accounts. You can discover this list from your own data. Destinations that most of your hosts contact on most days are, almost by definition, the services you run on.
- Known-neutral: popular public sites and content networks. Traffic to them is unremarkable in small amounts. A large upload to a consumer file-sharing or personal storage service is the classic gray area, legitimate for some users and not for others.
- Unknown: this could be a bare IP address with no DNS name behind it, or a name registered so recently it has no history. It could be a hosting provider you have no business with, or an address in a region your organization never deals with. None of these is proof of anything. All of them raise the score of a transfer that is already suspicious on volume, time, or role.
Keep the known-good list short and explicit, and review it when you refresh the baseline. The mistake to avoid is the opposite one: treating everything that is not on a blocklist as fine. Reputation is a way to weight a signal, never a way to dismiss one on its own.
Building the baseline from flow data
The raw material is a record of conversations with byte counts. The two everyday sources are flow exports from your routers or switches and session logs from your firewall. Both give you, per conversation, the internal host, the external destination, the protocol and port, the start time, and the bytes in each direction. The security case for collecting flows in the first place is made in Flow Data for Security. Here we assume you have some and want to turn it into rows. If part of your estate lives in a cloud provider, that provider's own egress metrics fill the same role for those hosts. Egress and Cost Visibility explains where to find them.
The recipe is a numbered sequence you can follow by hand for a dozen hosts or script for a few hundred:
- 1. Choose the measurement period. Two working weeks is the minimum; four is better, because it includes a month-end and at least one heavy day.
- 2. Filter to outbound. Keep only conversations whose destination is outside your address space. Internal traffic is a different problem.
- 3. Sum per host per day per window. For each internal host, add up bytes out for working hours, off hours, and weekends separately.
- 4. Count distinct destinations per host per day. Group by owning network where you can.
- 5. List the top destinations per host. The handful of addresses that carry most of its outbound bytes.
- 6. Record typical and high-water values. Use the median day as "typical" and the largest normal day as the high-water mark. Do this after throwing out any day you can explain as a one-off.
- 7. Compare with the role expectation. Any host whose measured profile contradicts its role table row gets investigated before its numbers are accepted as baseline.
Step seven is the one people skip, and it matters. If you baseline blindly, you baseline whatever is already happening, including anything that should not be. A workstation that "normally" uploads a few gigabytes a night is not normal; it is a question. The output of a scripted run looks roughly like this:
host role win typ_out hw_out dests/day top_dest (share) WS-FIN-04 workstation work 310 MB 1.1 GB 41 cloud-mail (52%) WS-FIN-04 workstation off 18 MB 95 MB 6 update-svc (80%) WS-FIN-04 workstation wkend 4 MB 22 MB 3 update-svc (90%) BACKUP-02 backup work 60 MB 400 MB 3 offsite-tgt (70%) BACKUP-02 backup off 170 GB 240 GB 2 offsite-tgt (99%) DB-01 database work 9 MB 30 MB 2 update-svc (95%) DB-01 database off 2 MB 6 MB 1 update-svc (100%) CAM-LOBBY-1 camera all 2.4 GB 2.6 GB 1 vendor-cloud (100%)
Even this small sample teaches something. The camera uploads more than the finance workstation every single day. That is completely fine, because it is a cloud camera and that is its job. Role first, numbers second.
A baseline table template per role
Here is the row layout to copy into whatever you keep documentation in. One row per baselined host; one row per segment for the crowd, with the per-device typical value noted. The example values are illustrative placeholders to show the shape, not targets.
| Host / segment | Role | Working window | Out / day (typ → high) | Off-hours out (typ → high) | Dests / day | Top destinations | Legit night jobs | Reviewed |
|---|---|---|---|---|---|---|---|---|
| Office VLAN (per device) | workstation | 08:00–18:00 local | fill in | fill in | fill in | cloud-mail, cloud-files, update-svc | patching only | date |
| WS-FIN-04 | workstation, finance | 07:30–17:00 local | fill in | fill in | fill in | cloud-mail, banking portal | none | date |
| FILES-01 | file server | always on | fill in (external only) | fill in | fill in | update-svc, offsite-sync | sync 23:00–01:00 | date |
| BACKUP-02 | backup target | 01:00–05:00 job | fill in | fill in | fill in | offsite-tgt | the backup itself | date |
| DB-01 | database | always on | fill in (expect near zero) | fill in | fill in | update-svc | none | date |
Three columns carry more weight than they look. Legit night jobs is the seed of the allowlist you will need for odd-hours alerting. Top destinations is the seed of your known-good list. And Reviewed is the date you last confirmed the row still matches reality, which brings us to the last question.
How often to refresh baselines
Baselines age. People change roles. Applications move to the cloud. A department starts using a new service. A server gets a new job without anyone updating the spreadsheet. A stale baseline generates false alarms, and false alarms get ignored, and ignored alarms are the same as no alarms. So refresh on a schedule:
- Segments and ordinary workstations: re-measure roughly every quarter, or after any change that touches the whole group. Examples are a new file-sharing service, a new endpoint agent, or a move to a different cloud email provider.
- Individually baselined servers: re-measure after every change ticket that touches them, and at least twice a year regardless.
- Sensitive hosts: glance at their rows monthly. It takes minutes and it is where the cost of drift is highest.
- After an incident: always. Whatever you learned changes what normal means.
Two rules keep refreshing honest. First, never refresh during an investigation. If a host is under suspicion, freeze its baseline until the case is closed, or you will quietly normalize the very behaviour you are investigating. Second, when a refresh moves a host's numbers a long way, treat that as a finding, not as a correction. "The finance workstation's off-hours upload baseline tripled since last quarter" is a sentence that deserves a reason before it deserves a new row. This is also your partial defence against the slow leak that predates your measurements. Every refresh is a chance to compare the new normal against the role expectation and ask, again, whether it makes sense.
Where to go next
You now have the artifact that everything else in this pillar depends on. It is a per-host and per-segment profile of outbound bytes, destinations, windows, and night jobs, checked against what each role should do. First-seen destinations are logged as events, and you have a home-made sense of which destinations are old friends. The next guide, Odd-Hours Transfers, turns the working-window column into your cheapest alert rule. Egress Monitoring on a Budget shows how to keep these numbers flowing with free tools. If you have not yet read it, What Data Exfiltration Looks Like explains why the comparison matters. The full series is at Spotting Data-Exfiltration Patterns.
Put this into practice with SysPeep Monitor
A baseline is only as good as the weeks of quiet collection behind it. SysPeep Monitor auto-discovers your devices and interfaces. It records traffic and health over SNMP and WMI around the clock. It graphs that data from one web console. So the working-window bump and the off-hours floor for each host are already on record when you sit down to fill in the table above. The free trial installs on any Windows machine.
Download Free TrialLearn About SysPeep