Detections a Small Team Can Actually Run
You have the logs flowing into one place. Now comes the question that stalls most small teams: what, exactly, are you looking for? Vendor rule libraries have hundreds of entries, most of which assume a staffed security operations center to sort through what they produce. You need the opposite — a short list of patterns that almost always mean something. They should fire rarely enough to be read every time. Each should come with a first response you can carry out yourself.
This guide is that list. Nine detections, each described from the defender's chair. Each entry covers what you look for in the logs, why it matters, and a sketch of the rule logic. It covers the legitimate cases that will trip the rule and how to tune them out. It also explains what to do first when the rule fires. None of them require a commercial platform. All of them can be run as a scheduled script or a saved search against the sources you already collect.
A word before the list. Every rule below has a threshold, and every threshold is a guess until you have watched your own network for a couple of weeks. The number that is right for a fifty-person office is wrong for a five-hundred-person one. Treat the numbers here as starting points, and treat the tuning notes as the actual instructions.
In this guide
- How each detection is written up
- Baselines: the number your thresholds are relative to
- The nine detections
- The detection table: severity and first response
- Running the rules without a platform
- Where to go next
How each detection is written up
Each entry follows the same shape, and the shape is worth copying when you write rules of your own. What you look for is the observation, in plain words. Why it matters is the reason a defender cares. The rule sketch names four things. First comes the source the rule reads. Next is the condition it matches. Then comes the threshold, expressed relative to your normal baseline rather than as a magic number. Finally, there is the confirmation window. This is how long the condition must hold, or how many occurrences must stack up, before it counts. Tuning lists the legitimate things that will trigger the rule so you can exclude them deliberately. First response is the one action to take before anything else.
Baselines: the number your thresholds are relative to
A baseline is simply what a measurement looks like on a boring day. How many failed logins does your directory see in an ordinary hour? How many files does a typical user open in ten minutes? What hours do your administrators actually work? You find these numbers by running the rule's query with no threshold for two weeks and looking at the distribution. Then you set the threshold well above the normal range. Do not set it at the highest value you ever saw. Set it far enough above the everyday one that a real event stands out. A threshold written as "five times the hourly average" survives growth and staff changes. A threshold written as "fifty" is correct until the day it is not.
The nine detections
1. Many failed logins, then a success, for one account
What you look for: a run of authentication failures for a single account, followed within a short window by a success for that same account. Pay special attention when the failures and the success come from the same source host or address.
Why it matters: people mistype a password two or three times and then get it right. Nobody mistypes it thirty times and then gets it right. A long run of failures ending in a success is the signature of guessing that worked. And the success is the part that turns a nuisance into an incident: someone is now inside with that account.
source: directory auth, VPN, RADIUS condition: failure count for one account, then a success for the same account threshold: failures well above the account's normal (start: 10+ in 10 minutes) window: success within 15 minutes of the last failure
Tuning: exclude the failure-then-success pattern where the failures come from a device with a cached old password. A phone after a password change is the classic case. Those failures show a steady trickle from one device, not a burst. Watch for lockout policies that stop the run early. A burst that ends in a lockout rather than a success is a lower-severity cousin worth its own alert.
First response: confirm with the account owner whether the success was them. If not, reset the credential and look at what the account did next. The walkthrough in correlation without a SIEM starts from exactly this event.
2. Failures spread wide rather than deep
What you look for: the inverse shape of the previous rule. Instead of many failures on one account, you see a few failures each across many accounts. Or you see one account failing from many different hosts. All this happens within a short window and often from a single source. Each individual account sees only one or two failures — too few to trigger a lockout or the rule above.
Why it matters: real users fail on their own account from their own machine. A pattern that touches dozens of accounts lightly is not a person forgetting a password. It is something systematically testing credentials, one or two tries per account. That shape stays below what rule 1 watches for, which is why you need rule 2.
source: directory auth, VPN, RADIUS condition: distinct accounts with at least one failure, grouped by source host threshold: distinct-account count well above normal (start: 10+ accounts) window: 30 minutes
Tuning: vulnerability scanners and some monitoring tools authenticate against many hosts on a schedule. Exclude them by source address and confirm the schedule matches. A single application server that holds credentials for many users can also trip this after a password policy change.
First response: identify the source host. If it is internal, isolate it from the network before anything else. If it is an external address, block it at the firewall and check whether any of the touched accounts recorded a success.
3. A new administrative account, or a user added to an admin group
What you look for: any account-created event where the new account is immediately placed in a privileged group. Also watch for any group-membership change that adds a member to a group granting administrative rights. Check the directory, individual servers, and the firewall or other network gear that keeps its own admin list.
Why it matters: this is the moment access is escalated. On a small network, legitimate additions to administrative groups happen a handful of times a year and always have a reason someone can state. An addition nobody can explain is the single highest-value alert on this list.
source: directory admin events, server local group events, device admin lists condition: member added to any group on your privileged-groups list threshold: one event (no baseline needed) window: immediate
Tuning: almost none. Maintain the list of privileged groups carefully, and require that every legitimate addition has a ticket number you can match against the alert. Your onboarding process may create accounts through a script. If so, exclude the script's service account only if the script itself cannot add members to privileged groups.
First response: match the event to a change request. If there is none, remove the membership and disable the account if it is new. If no change request exists, treat the account that made the change as the subject of the investigation.
4. Administrative logins outside working hours
What you look for: an interactive or remote-desktop logon by any member of a privileged group at a time when your administrators are not normally working. That means overnight, on weekends, or on holidays your organization observes.
Why it matters: administrators do sometimes work at odd hours, but they know when they did it. Someone using an administrator's credentials without permission tends to prefer the hours when nobody is watching. The rule turns "nobody was watching" into "the log was watching."
source: directory auth (interactive and remote logon types) condition: account is in a privileged group AND logon time outside the schedule threshold: one event window: immediate; optionally suppress during declared maintenance windows
Tuning: keep a simple calendar of declared maintenance windows and suppress the rule during them. Distinguish interactive logons from the network logons that backup jobs and monitoring make on behalf of service accounts. Those network logons should be excluded here and caught by rule 6 if they become interactive.
First response: ask the administrator, directly and immediately. If it was them, note it and move on. If it was not, reset the credential and pull everything that session touched.
5. Endpoint protection disabled on a host
What you look for: a status event from the endpoint protection console. It may say real-time protection was turned off, the service was stopped, or tamper protection was disabled. Or it may say a host that was reporting regularly has gone quiet while it is otherwise online.
Why it matters: protection does not switch itself off. Either an administrator did it — and that belongs in a ticket — or someone with local administrator rights on that host did it. The latter is very commonly the step right before something is installed that protection would have stopped.
source: endpoint protection console condition: protection-disabled or agent-silent event for a host threshold: one event; for "silent," no check-in for 2x the normal interval window: immediate; agent-silent confirmed by a successful ping to the host
Tuning: exclude hosts in a documented maintenance state. Confirm that the "agent silent" variant does not fire for machines that are simply powered off. Pair it with a reachability check so that only online but silent hosts count.
First response: re-enable protection remotely if you can, then find out who disabled it and why. If the host was silent while online, treat it as compromised until a scan says otherwise.
6. A service account logging in interactively
What you look for: an account that exists to run a scheduled job, a backup, or an application. This is a service account. Watch for it performing an interactive or remote-desktop logon, which is something a service never needs to do.
Why it matters: service accounts are attractive because they often have broad rights and rarely have their passwords changed. They are not associated with a person who would notice something odd. A service account sitting at a keyboard, or opening a remote desktop session, means a person is using it. And that person is not supposed to exist.
source: directory auth condition: account on the service-account list AND logon type is interactive or remote threshold: one event window: immediate
Tuning: the rule is only as good as your service-account list. Keep it as a group in the directory or a file the script reads. Review it whenever an application is installed or retired. Some vendors' installers log on interactively once during setup. Note those as expected during the install ticket.
First response: identify the source host of the logon and isolate it. Reset the service account's credential during a window when the job it runs can tolerate a restart.
7. Mass file access from one account in a short window
What you look for: a single account opening, reading, renaming, or deleting far more files on a file share than any user does in normal work. This happens in minutes rather than a day. It is visible in file-server audit logs if object access auditing is enabled on the shares that matter.
Why it matters: people work on a few files at a time. Software that encrypts a share, or a person copying a department's entire folder to a removable drive, touches thousands. The burst is the tell. Catching it in the first minutes is the difference between restoring one folder and restoring a server.
source: file server object-access audit log condition: distinct files touched by one account threshold: well above the busiest normal user (start: 5x the hourly maximum) window: 10 minutes, then alert once per account per hour
Tuning: backup agents, indexing services, and antivirus scans touch everything by design; exclude their accounts explicitly. Migrations and bulk archive jobs should be announced and suppressed by ticket. Enable auditing only on the shares that hold data you would miss, or the volume will overwhelm the rule.
First response: disconnect that account's sessions from the file server and disable the account. Then check whether the touched files still open normally. Involve backups early.
8. A device that stops sending logs
What you look for: a source on your inventory that has sent nothing for longer than its normal quiet period. The device itself is still reachable. Not a security event in the log — the absence of events.
Why it matters: every other rule on this list assumes the logs are arriving. A firewall may have had its syslog switched off. A domain controller's forwarding may have broken after a patch. A collector disk may have filled up. Each one is a blind spot. And a blind spot that opens right before an incident is rarely a coincidence. This is the rule that watches the watchers.
source: the collector itself (last-seen time per source) condition: no events from a source for longer than its expected gap threshold: 2x the source's longest normal quiet interval window: confirm with a reachability check; alert if online and silent
Tuning: quiet sources like endpoint protection legitimately go hours without speaking. Set their expected gap accordingly, or send them a daily heartbeat event to remove the ambiguity. Loud sources like the firewall should never be silent for more than a few minutes.
First response: check the device's log configuration and the collector's disk before assuming anything sinister. Then, if the configuration was changed, find out by whom.
9. A VPN login from a place the user has never used
What you look for: a successful VPN or remote-access login for a user from a source network, region, or country absent from that user's history. Also watch for two successful logins for the same user from locations too far apart for the time between them.
Why it matters: a valid password used from an unfamiliar place is the clearest sign that the password is no longer private. Users travel, so "new location" alone is not proof. But a new location that has never appeared before deserves a phone call every time. So does one that is physically impossible given the previous login.
source: VPN / remote-access logs, with source addresses mapped to a region condition: successful login from a region absent from the user's history threshold: one event; "impossible travel" if two regions within a short gap window: immediate; history window of 90 days per user
Tuning: mobile carriers and privacy-focused browsers can make a user appear to move. A short allowance for known carrier ranges reduces noise. Build the per-user history from the logs you already retain before switching the rule on. That keeps it from firing for everyone on day one.
First response: contact the user through a channel other than email. If they did not log in, terminate the session, reset the credential, and review what the session touched.
The detection table: severity and first response
| # | Detection | Severity | Expected rate | First response |
|---|---|---|---|---|
| 1 | Failures then success, one account | High | Rare | Confirm with owner; reset if not them; trace what the account did next |
| 2 | Failures spread across many accounts | High | Rare | Identify the source; isolate or block it; check for any success |
| 3 | New admin or added to admin group | Critical | A few per year | Match to a ticket; remove and investigate the actor if none |
| 4 | Admin login outside hours | Medium | Occasional | Ask the admin directly; reset and trace if not them |
| 5 | Endpoint protection disabled | High | Rare | Re-enable; find who and why; treat silent-while-online as compromised |
| 6 | Service account interactive logon | High | Very rare | Isolate the source host; rotate the credential in a safe window |
| 7 | Mass file access, one account | Critical | Very rare | Disconnect and disable the account; check files; involve backups |
| 8 | Source stopped sending logs | Medium | Occasional | Check config and collector disk; find who changed it |
| 9 | VPN login from a never-seen place | High | Occasional | Contact user out of band; terminate session and reset if not them |
Tip: "expected rate" is the tuning target. If a rule marked "rare" fires daily, the rule is wrong for your network, not the network wrong for the rule. Adjust the threshold or add an exclusion. But write down what you excluded and why. An exclusion nobody remembers is a hole nobody is watching.
Running the rules without a platform
None of this needs a correlation engine. Rules 3, 5, 6, and 8 are single-event matches. A scheduled script that reads the last hour of the relevant source, matches a pattern, and sends an email will do. Rules 1, 2, 7, and 9 need a count or a comparison over a window. That is a slightly longer script: group by account or source, count, compare to a threshold. It is still well within what a shell or scripting language does in a few dozen lines. Rule 4 is a single event plus a lookup against a schedule. Run the batch every fifteen minutes for the high-severity rules and hourly for the rest. The mechanics of turning a match into a notification that reaches you are covered in alerting on logs.
Start with three rules, not nine: the new-admin rule, the protection-disabled rule, and the failures-then-success rule. Tune them until they are quiet on a normal week. Then add the next three. A small set you trust is worth more than a full set you have learned to ignore.
Where to go next
When one of these rules fires, the next question is always "and then what did that account or host do?" That is correlation. The following guide in Security Log Monitoring on a Budget (SIEM-lite) shows how to answer it by hand and by script. It uses the same sources these rules read. If you have not yet chosen those sources, the logs that matter most is where to start.
Put this into practice with SysPeep Monitor
Rule 8 — the source that stops sending — is really a monitoring problem. The surest way to catch it is to watch the machines themselves. SysPeep Monitor tracks the domain controllers, firewall, file servers, and collector behind these detections over SNMP and WMI. It alerts when one goes down or its disk fills. It keeps the uptime history that tells you whether a quiet source is off or just offline. The free trial installs on any Windows machine in minutes.
Download Free TrialLearn About SysPeep