Odd-Hours Transfers: Alerting on Time as a Signal
Of all the ways to notice data leaving your network, the clock is the cheapest. It needs no reputation feed, no machine learning, and barely any math. It rests on one observation that is almost embarrassingly simple: people work during working hours, and their computers mostly do too. A workstation pushing a large amount of data to the internet at three in the morning is unusual by definition. That is because the person who uses it is asleep and desktops do not have night jobs of their own.
The catch is that "unusual by definition" is not the same as "worth waking someone up." Networks are surprisingly busy at night. Patches download, backups run, cloud sync finishes what it started at five, and somebody in another time zone is always working. An after-hours rule written carelessly fires every night and is muted within a week. Written carefully, the same rule stays quiet for months and then, one night, tells you exactly the thing you needed to know.
This guide is about writing it carefully. You will build the allowlist of legitimate night jobs first, then the rule itself. Then you will build the multi-condition version that keeps noise down, followed by the handling for time zones and remote workers. Finally, you will build the alert text that actually helps a tired human at three in the morning. This is the defender's side only: what you watch and what you do when the watch fires.
In this guide
- Why time is the cheapest signal
- Step one: the allowlist of legitimate night jobs
- The basic rule: bytes out, outside the window, above a floor
- Multi-condition alerts that stay quiet
- Time zones, shift workers, and remote workers
- The alert text that helps at 3 a.m.
- Tuning in the first two weeks
- When it fires
- Where to go next
Why time is the cheapest signal
Most exfiltration signals require you to know a lot about a host before they mean anything. Volume needs a baseline. A new destination needs a history of old ones. Role mismatch needs an inventory that says what each device is for. Time needs one fact: when is this host supposed to be busy? For a workstation, that is the working hours of the person who sits at it. For a server, it is the schedule of the jobs it runs. Both are easy to find out and slow to change.
Time is also a signal that people who steal data have trouble avoiding. Moving a large amount of data during the day, while the user is present and the network is busy, is possible. Some cases do exactly that, hiding inside daytime bulk. But night is tempting precisely because nobody is watching. An after-hours rule takes that temptation away, or at least makes it expensive. The point is not that theft only happens at night. It is that theft at night is nearly free to detect, so you should collect that free detection first.
The honest limit: time on its own is weak. A "workstation active at night" alert with no other conditions will fire on patch downloads and on the person who works late. It will fire on a laptop left running a sync, and on every misconfigured scheduled task in the building. So the rule you will actually run says three things at once: outside the window, outbound, and above a size floor. Everything in this guide is about getting those three qualifiers right, then adding more.
Step one: the allowlist of legitimate night jobs
Do not write the rule first. Write the allowlist first: the list of things that are supposed to move data out of your network after hours. If you do it the other way round, the rule will fire on your own backups for a week. In that case, you will add exceptions in a hurry, and the exceptions will be sloppy. Sloppy exceptions are how a real event slips through. "Ignore anything from the backup server at night" quietly also ignores the backup server being used to push data somewhere it has never gone.
A good allowlist entry is specific on four points: source (which host), destination (which external address, name, or owning network), window (which hours and days), and expected size (roughly how much). Add an owner — the person who can confirm the job still exists — and a review date. Anything that matches all four points is expected. Anything that matches only some of them is exactly what you want to hear about. That could be the backup server sending at the right time to the wrong place. Or it could be sending to the right place at the wrong time, or ten times the usual amount.
To discover the entries, run an after-hours outbound report for two or three weeks before you alert on anything. Include every host that sent more than a modest amount outside working hours, with destination and volume. Then explain every line. Most will be obvious. A few will need a phone call. One or two, in most organizations, will turn out to be forgotten jobs nobody owns anymore — those get removed rather than allowlisted. Here is what the finished list looks like:
| Source | Destination | Window | Expected size | Owner | Reviewed |
|---|---|---|---|---|---|
| BACKUP-02 | off-site backup target (owning network) | 01:00–05:00, nightly | tens to hundreds of GB | backup admin | date |
| FILES-01 | cloud sync target | 23:00–01:00, weekdays | single-digit GB | storage admin | date |
| Office VLAN (any) | OS and application update services | any | small per host; downloads dominate | desktop team | date |
| ERP-APP-01 | partner data exchange host | Fri 22:00–23:00 | hundreds of MB | finance systems | date |
| CAM-* (camera VLAN) | camera vendor cloud | any | steady, a few GB per camera per day | facilities | date |
| MAIL-RELAY-01 | any (recipient mail servers) | any | small, steady | messaging admin | date |
| MON-01 | notification and status services | any | tiny | you | date |
Notice that the workstation entry is deliberately narrow: updates only, and downloads dominate. Workstations do not get a blanket night exemption, ever. That single decision is what makes the rest of the rule work.
The basic rule: bytes out, outside the window, above a floor
With the allowlist in hand, the core rule is short. For each host, define a working window from the baseline. For workstations, it is typically the user's hours plus an hour on either side. For servers with a separate job schedule, it is typically "always". Then, for each evaluation period outside that window, add up the host's outbound bytes to destinations that are not covered by an allowlist entry. If the total exceeds a size floor, raise a notice.
The size floor is the part that needs thought. Its job is to absorb the ordinary background chatter that every host produces at night — sync leftovers, telemetry, the small upstream half of a download. Set it from the host's own off-hours baseline. Take the highest off-hours total you saw in a normal week and add a comfortable margin. That way a night twice as chatty as the chattiest normal night still stays quiet. For a typical office workstation that lands somewhere in the hundreds of megabytes. For a media workstation, it is higher. For a database server, it is very low, because the server should have almost no night story at all. Different floors per role, not one number for everyone.
Choose the evaluation period deliberately too. Hourly buckets catch a fast burst quickly. A per-night total catches a slow trickle that never exceeds the floor in any single hour. Run both if you can; if you can only run one, run the per-night total, because it catches both shapes, just later. Here is the rule as a sketch, in no particular product's syntax:
RULE after_hours_upload
FOR EACH host IN baselined_hosts
window = host.working_window # e.g. Mon-Fri 07:00-19:00 local
floor = host.offhours_highwater * 2 # per-role, from the baseline
period = last 60 minutes AND ALSO since window_end (per-night)
IF now NOT IN window
out_bytes = SUM(bytes_out)
WHERE src = host
AND dst NOT IN internal_ranges
AND (host, dst, now) NOT MATCHED BY allowlist
IF out_bytes > floor
RAISE notice(host, out_bytes, top_dst, first_seen(top_dst), floor)
Two details in the sketch matter more than they look. The allowlist is matched on the combination of host, destination, and time — not on the host alone. That is what keeps the backup server honest. And the raise carries first_seen for the top destination, because that single flag is what turns a notice into something more.
Multi-condition alerts that stay quiet
The basic rule produces notices: events written to a log or a daily report, not pages. To decide what deserves a page, combine time with the other exfiltration dimensions. Three conditions do most of the work:
- Odd hour — the basic rule fired: outbound, outside the window, above the floor, not allowlisted.
- New destination — the destination that received most of those bytes is first-seen for this host, or first-seen for the whole network (stronger).
- Role mismatch — the host's role does not include sending data out: a workstation, a printer, a camera doing something other than its vendor stream, a database server doing anything at all.
Then tier the response by how many conditions line up. One condition is a notice: it goes into the morning report and nobody is woken. Two is a warning: someone looks at it first thing, and if the host can reach sensitive data, someone looks at it now. Three is a page: whoever is on call gets it immediately, with the alert text described below. The tiers are not a scoring system with false precision. They are a way of making sure that the alert which wakes someone up is one where every innocent explanation has already been made unlikely. The reasoning behind that — why single signals mislead and combinations convince — is laid out in What Data Exfiltration Looks Like.
You can add a fourth condition if your data supports it: unusual protocol. That means the transfer used a port or protocol this host has never used before. And you can add a modifier rather than a condition: sensitivity. A host that can reach the finance share or the customer database gets its thresholds lowered one tier, so two conditions page instead of warn. The cost of a false page on that host is one grumpy person. The cost of a missed one is the reason you are reading this.
Time zones, shift workers, and remote workers
"Outside working hours" is only meaningful if you know whose working hours. This is where naive after-hours rules go wrong, so handle it explicitly.
Store the window per host, not per organization
The working window is a property of the host (or its group), recorded in the baseline, not a single company-wide setting. Most hosts inherit a group window — "office VLAN, weekdays, seven to seven local" — and a minority get their own. The night-shift operators' PCs get a night window. The support desk that covers another continent gets that continent's window. A server's window is its job schedule, which is why servers mostly live on the allowlist rather than in the rule.
Keep logs in one time reference and rules in local time
Flow records and firewall logs should be timestamped consistently, ideally in a single reference such as UTC, so that evidence from different devices lines up. But the rule should evaluate in the host's local time, because "three in the morning" only means something relative to the person. Convert once, in the rule, and never by hand. A rule that is off by the daylight-saving shift will spend half the year firing an hour early and the other half missing an hour of real night.
Remote and travelling workers
A laptop that connects through your VPN from three time zones away is not on your schedule. Give remote users a wider window, or a window tied to their home time zone, and lean harder on the other conditions for them. Their destination history still applies. A remote worker's laptop talking to the company file service at their local nine in the morning is fine. The same laptop uploading gigabytes to a first-seen address at anyone's three in the morning is not.
Holidays and leave
A public holiday makes the whole office "off hours" for a day. That is the single best day to run the per-night version of the rule against the daytime as well. If your HR or calendar system can tell you that a specific user is on leave, that is gold. A workstation active while its user is on a beach is a strong signal on its own, and it costs nothing but a lookup.
The alert text that helps at 3 a.m.
An alert that says "after-hours upload threshold exceeded on 10.20.4.71" is, at three in the morning, nearly useless. The person receiving it has to look up the address, find the owner, open the flow tool, work out what happened, and decide what to do. They do all this while half asleep. Every one of those lookups can be done by the rule before it sends anything. The alert should carry:
- Host, role, and location — the name, what it is for, and where it sits. "WS-FIN-04, finance workstation, 3rd floor" beats an address.
- User, if known — the assigned user, and their leave status if your directory can supply it.
- Destination — the address, the DNS name if one was looked up, the owning network, and the first-seen flag: new for this host, new for everyone, or known.
- Volume against baseline — bytes out in the window, the host's off-hours high-water mark, and the ratio between them. "3.8 GB against a normal night of 60 MB" is instantly meaningful.
- Time span and protocol — when it started, whether it is still running, and what port or protocol carried it, with a note if that is new for the host.
- Which conditions matched — odd hour, new destination, role mismatch, unusual protocol — so the reader knows why this is a page and not a notice.
- A link to the evidence — the exact flow or firewall query that produced the numbers, so the reader can see the raw records in one click.
- The first agreed action — not "investigate," but the concrete step your playbook says comes first.
Put together, the page reads something like this:
PAGE after-hours upload | 3 of 4 conditions | still active Host: WS-FIN-04 (finance workstation, 3rd floor) User: j.doe -- directory shows ON LEAVE this week Destination: 203.0.113.58 (no DNS name recorded; hosting provider; NEW for whole network) Volume: 3.8 GB out since 02:41 local | normal night: 60 MB | ~60x Protocol: HTTPS (normal for host) -- direct to IP, no lookup seen Matched: odd hour, new destination, role mismatch Evidence: flow query saved as case-tmp-0417 (do not modify) First step: preserve the flow export, then call on-call lead before any block
And two things to leave out. Do not include a verdict ("possible data theft by user"). The alert reports what the network saw, and intent is not the network's to judge. And do not include instructions that only a specialist could follow. The person on call is you or someone like you, and the alert should make sense to them.
Tuning in the first two weeks
Expect noise at first, and treat every false alarm as information. Each one has exactly one of three causes. It might be a missing allowlist entry (add it, properly, with all four points). It might be a floor set too low for that role (raise it, from data, not from irritation). Or it might be a wrong working window for that host or group (fix the window). What you must never do is widen the window or raise the floor globally just to make the noise stop. That is how the rule ends up unable to see the very thing it was built for.
Tip: keep a count of notices, warnings, and pages per week, and write it down. In a healthy setup the notices settle into a low, steady number, warnings are rare, and pages are rarer still. If notices climb without a matching change in the business, something has changed on the network — and that is itself worth a look before you touch the rule.
One more tuning trap: the allowlist can rot in the other direction. A job that was decommissioned still has an entry, and that entry now silently exempts whatever else that host does at night. The review date on every entry exists to catch this. When the date passes, the owner confirms the job still runs, or the entry goes.
When it fires
A page at three in the morning creates a strong urge to do something: block the address, pull the cable, disable the account. Sometimes that is right. But the decision about containment depends on what the host can reach, whether the transfer is still running, and who is allowed to make that call. And those should be settled in daylight, in advance, not improvised at night. Write the first agreed action into the alert, as above, and make sure the on-call person knows who to phone. The full sequence — confirm, identify, preserve evidence, assess sensitivity, contain proportionately, escalate, communicate — is the subject of the investigation playbook in this pillar. The only rule that applies before any of that: preserve the flow records first. They are cheap to export and impossible to recreate.
Where to go next
You now have the cheapest exfiltration detector there is, built the right way round. The allowlist comes first, then a per-host rule with a per-role floor. Then come tiers that combine time with destination and role so that the alert which wakes someone up has earned it. And there is alert text that a tired human can act on. The working windows and floors come from the profiles described in Outbound Baselines. The data that feeds them is the subject of Egress Monitoring on a Budget. The whole series lives at Spotting Data-Exfiltration Patterns.
Put this into practice with SysPeep Monitor
A time-based rule is only as trustworthy as the alert pipeline underneath it. SysPeep Monitor lets you build threshold and alert rules with dependencies designed to keep noise down. It delivers them by email and other notification channels, and shows the interface traffic behind each alert on a live dashboard in the same web console. So the traffic graph behind a 3 a.m. alert is already waiting for you. The free trial installs on any Windows machine.
Download Free TrialLearn About SysPeep