Triaging IDS Alerts Without a SOC
You turned the sensor on Friday afternoon. On Monday morning there are several thousand alerts waiting, with names like "EXPLOIT attempted overflow," "MALWARE possible check-in," and "SCAN behavioral unusual port." Some are marked high priority. You have no security operations center, no analyst, and no idea which of these — if any — means something is actually wrong. The temptation is to close the tab and decide that intrusion detection is for bigger companies.
Resist it, because the skill you are missing is smaller than it looks. Triage is not the ability to recognize every attack. It is the habit of asking the same five questions about an alert and answering them from data you already have. Then you turn the answer into one of four decisions. Do that for a couple of weeks and the several thousand becomes a few dozen. The few dozen becomes a handful, and the handful is worth your time. This guide gives you the questions, the data sources, and the decision flow. It gives you the tuning moves that silence what should never have fired. It gives you a daily and weekly rhythm that fits around a real job, and a worked example from raw alert to decision.
In this guide
- What a fresh sensor really looks like
- Severity is a hint, not a verdict
- The five context questions
- Enrichment: answering the questions from your own data
- The triage flow: dismiss, watch, investigate, escalate
- Tuning noisy signatures
- The daily fifteen minutes and the weekly review
- A worked example: one alert, start to finish
- Where to go next
What a fresh sensor really looks like
A community rule set is written for everyone, which means most of it is written for someone else. It contains rules for web frameworks you do not run and database servers you do not have. It contains rules for industrial protocols you have never heard of, and malware families that target platforms not present on your network. All of those rules are enabled by default, because the rule authors cannot know what you run. A public-facing server receives the ordinary background noise of the internet — automated probes trying every known weakness against every address they can find. When it does, a large fraction of those rules fire, correctly, on attempts that could never have succeeded.
Add your own network's habits. There is a vulnerability scanner that looks exactly like an attacker because it is imitating one. There is a backup product with an odd protocol. There is a monitoring system that polls everything constantly. And there are users' browsers making requests that resemble old exploit patterns. None of this is a failure of the sensor. It is what a sensor with no knowledge of your network produces. The whole job of triage and tuning is to give it that knowledge, one decision at a time.
So the first rule of a fresh sensor: do not read alerts one by one. Sort by rule and count. The top ten rules by volume are usually most of the noise, and each one is a single tuning decision away from silence. Only after that does reading individual alerts become worthwhile.
Severity is a hint, not a verdict
Every alert carries a priority or severity, assigned by the rule's author. It is tempting to sort by it and work from the top. Do not, or at least not only. The author assigned that number without knowing whether the target exists on your network or whether it runs the affected software. The author did not know whether the connection succeeded or whether the source is your own scanner. A "high" alert for an exploit against software you do not run is noise with a scary label. A "low" alert for a policy violation — a cleartext login from a server — may be the most important thing the sensor found all week. It describes a real weakness in your environment rather than a failed attempt against it.
Use severity as a tie-breaker between alerts that have passed the context questions, not as the reason to look at one in the first place. What matters is not how bad the rule's author thought the pattern was; it is how much the pattern could matter here.
The five context questions
These are the questions that turn an alert into a decision. Ask them in order; most alerts fall out at the first or second.
1. Is the target actually running the thing the signature is about?
An exploit attempt against a web application framework is meaningless if the target is a printer, a Windows file server, or a web server running a different framework. Most inbound exploit alerts fail this question, and the answer comes from your inventory. If you cannot answer it, that is a finding about your inventory, not the sensor.
2. Did the connection complete?
A probe that was rejected by the firewall, answered with a reset, or never answered at all is an attempt with no consequence. A signature can match on the first packet of a connection that went nowhere. Flow records tell you whether the target answered and how many bytes went each way. They come from the sensor's own metadata, from your firewall, or from a flow collector. No reply, no problem, most of the time.
3. Did the target talk back, or out, afterward?
This is the question that separates a failed attempt from a successful one. Suppose a server received an exploit attempt. Then, in the following minutes, it made a new outbound connection it has never made before. Or, in those minutes, it resolved a name it has never resolved, or started talking to another internal host on a new port. In any of those cases, the attempt may have worked. If the server went on doing exactly what it always does, it probably did not. Answering this needs connection history for the target. That is why a sensor's metadata logs and your monitoring's traffic history matter more than any single alert.
4. Is the source internal?
An outside address probing your public web server is Tuesday. An inside address doing the same thing is a different category entirely. It might be an infected machine, a curious employee, an unauthorized scanner, or a device that should not be on your network. Any alert with an internal source deserves at least a look. An internal source scanning or attempting exploits against other internal hosts deserves it today. The internal scanning patterns guide covers how to distinguish a scan from a chatty application and what to do when it is real.
5. Has this fired before?
Suppose a rule has fired daily for a month against the same target from a rotating set of outside addresses. Nothing interesting followed those alerts. That rule is background noise and a tuning candidate. A rule that has never fired before, or has never fired from this source or against this target, is new information. History is what lets you tell them apart. A sensor that has been running for a while is far more useful than one that started yesterday. That is because "has this fired before" has a real answer.
Enrichment: answering the questions from your own data
You do not need a security platform to answer the five questions. You need four sources of data that most small shops already have or can turn on quickly.
| Data source | Answers which questions | Where a small shop gets it |
|---|---|---|
| Inventory: what each host is and what it runs | 1 (running the thing?), 4 (source internal?) | Your monitoring system's discovered device list, a spreadsheet, the asset tool — whichever is most current |
| Flow and connection records | 2 (completed?), 3 (talked out afterward?) | The sensor's own metadata logs; firewall connection logs; a flow collector if you run one |
| DNS query logs | 3 (talked out afterward?), plus what a host was looking for | The sensor's DNS log; your internal resolver's query log |
| Monitoring history: CPU, traffic, process counts over time | 3 (behaving differently?), 5 (seen before?) | The graphs your monitoring already draws for every server |
The enrichment habit is simple. When an alert survives the first question, open the target host in your monitoring and look at the hour around the alert timestamp. Did its outbound traffic change shape? Did its CPU do anything unusual? Then look at its connection records for that hour. That two-minute look answers questions two and three for most alerts, and the answer is usually "nothing happened."
Tip: make the enrichment fast by making it a fixed route. Keep a bookmark to the host page in your monitoring. Keep a saved search in your log tool for "this source, this hour," and the inventory open in a tab. Triage that takes eight clicks per alert will not survive the week; triage that takes two might.
The triage flow: dismiss, watch, investigate, escalate
Every alert that reaches a human ends in one of four decisions. Naming them keeps you from dithering, and recording them builds the history that question five depends on.
- Dismiss. The alert cannot matter here: wrong target, connection never completed, known-good source. Close it, and if the same reasoning will apply next time, tune the rule so it does not come back.
- Watch. The alert could matter but nothing followed it. Note the source, target, and rule; check again at the next daily review to see whether anything has developed. Most watched alerts expire quietly after a day or two.
- Investigate. Something followed the alert — new outbound behavior, an internal source, a first-time pattern. Spend real time: connection history, DNS, what the host is doing now. This is where the sensor's metadata and your monitoring earn their keep.
- Escalate. Investigation found evidence of a real compromise or an active policy breach that needs someone else's decision. Isolate the host, notify a manager, and engage outside help. Your incident process takes over from here.
| Situation | Decision | Why |
|---|---|---|
| Inbound exploit attempt; target does not run the affected software | Dismiss, then disable or narrow the rule | Cannot succeed here; the sensor cannot know that but you do |
| Inbound exploit attempt; target runs it; no reply from the target | Dismiss, note the source | Attempt with no consequence; the firewall or the service refused it |
| Inbound exploit attempt; target runs it; connection completed; nothing changed afterward | Watch | Probably a failed attempt; confirm tomorrow that nothing developed |
| Inbound attempt; connection completed; target made new outbound connections afterward | Investigate now | The pattern of a successful exploit followed by a call home |
| Any scan or exploit alert with an internal source not on your allowlist | Investigate today | Infected host, unauthorized tool, or rogue device |
| Known malware check-in from an internal host, any confidence | Investigate now; escalate if confirmed | Even a false positive here deserves ten minutes; a true one is an incident |
| Policy rule: cleartext credential or forbidden protocol from a server | Investigate this week; fix the cause | Not an attack, but a real weakness you now know about |
| Rule that fires constantly on your own scanner or monitoring host | Dismiss and suppress by source | Known-good behavior that will never stop |
Tuning noisy signatures
Triage decides what to do with an alert; tuning decides whether it should have existed. Signature engines give you three tools, from bluntest to most precise.
Disable rules that cannot apply to you. If you run no instance of a given database product, every rule about attacks on it is pure noise. Rule sets are organized into categories and every rule carries a unique ID. So you can disable by category ("nothing about industrial control protocols") or by individual rule. Keep the list of disabled rules in a file with a one-line reason for each. In a year you will not remember why. A rule disabled for a server you have since deployed is a real gap.
Threshold rules that fire on benign volume. Some rules are correct but chatty. Examples are a scan-detection rule that fires once per probed port, and an alert on a protocol that your monitoring uses every minute. A threshold tells the engine to alert at most once per source per interval, or only after a count is reached. The rule stays on; the volume drops to something a human can read.
Suppress known-good sources. Your vulnerability scanner, your monitoring server, the backup host: for specific rules, from specific addresses, you never want to hear about it. Suppression silences a rule for a source or destination while leaving it active for everyone else. Be stingy with this; suppressing a rule for a whole subnet because one host was noisy is how real alerts disappear.
In the configuration format the Snort and Suricata family shares, those last two look like this:
# Vulnerability scanner: never alert on this scan rule from that host suppress gen_id 1, sig_id 9000301, track by_src, ip 10.10.0.5 # Chatty policy rule: at most one alert per source per hour threshold gen_id 1, sig_id 9000302, type limit, track by_src, count 1, seconds 3600
Two guardrails. First, tune only after triage. Disable a rule because you have looked at what it caught and found nothing that could matter, not because its count was high. Second, keep a rule in "watch" for a week before silencing it entirely. Some noisy rules occasionally catch something real, and a threshold preserves that while removing the flood. The anomaly-detection side of this problem has its own tuning discipline. It involves behavioral alerts that start noisy because they have not learned your network yet. That discipline is covered in Tuning Out Anomaly Alert False Positives. The moves here are specific to signatures, where the question is not "what is normal" but "does this rule describe something that could happen here." And if you find yourself dismissing alerts without reading them, stop and run the self-audit in Diagnosing Alert Fatigue. Dismissing alerts without reading them is exactly how the important one gets missed.
The daily fifteen minutes and the weekly review
Triage without a SOC is a habit, not a shift. Two fixed routines carry it.
Daily, fifteen minutes, same time every day. Open the sensor's alerts for the last day sorted by rule and count. Scan the list for anything new. Look for a rule you have not seen fire before, a source or target you do not recognize, or an internal source on a scan or exploit rule. Apply the five questions to those only. Record the decision for anything that reached "watch" or beyond. Check yesterday's watched items. Close the tab. If fifteen minutes is not enough, the sensor needs tuning, not more of your time.
Weekly, an hour, on a calendar. Look at the week's alert counts by rule. The top five are your tuning candidates; take each through triage and decide whether to disable, threshold, suppress, or keep. Review the watch list and expire what went nowhere. Confirm the rule-update job ran and skim what it added. Note anything the sensor revealed about the network itself — a forgotten service, a wrong firewall rule — and hand it to whoever fixes such things. Write one line in a log: alerts this week, decisions made, rules changed. That log is the history that makes next month's triage faster.
A worked example: one alert, start to finish
Monday morning, the daily fifteen minutes. Among the usual background, one alert stands out because the rule has not fired before:
[1:9000410:1] EXPLOIT Attempted command injection in web request [Priority: 1]
{TCP} 198.51.100.77:40318 -> 10.10.0.30:443
in_iface: edge-inside flow: established, to_server http.status: 404
Priority one, exploit category, aimed at an internal server from an outside address. Here is the walk through the questions.
- Running the thing? Inventory says 10.10.0.30 is the public web server. It does run a web application, so the rule is relevant. Question one passes, which is already unusual.
- Connection completed? The alert's own flow fields say the session was established, and the web server answered with a "not found" status. So the request reached the application and was rejected. Completed, but refused.
- Talked out afterward? Open the host in monitoring: outbound traffic for the following hour is flat and matches every other Monday. The sensor's connection log shows the server's only outbound sessions were to the update mirror it always uses and the monitoring server polling it. The DNS log shows no new names. Nothing followed.
- Source internal? No. An outside address with no prior history.
- Fired before? Not this rule. But the same source address appears in a dozen other dismissed alerts over the weekend. They are all against the same server, all different exploit rules, all answered with errors. That is the fingerprint of an automated scanner trying its whole list.
Decision: watch. The attempt was real and the target was relevant, but it failed and nothing developed. Note the source, check the host again tomorrow, and if the outside address keeps at it, ask the firewall to stop entertaining it. Tuning: keep the rule, because a request that reaches the application matters. Add a threshold so a scanner running the full list produces one alert per hour instead of a flood. Finding for the network: the web server answered a request that should have been filtered by the firewall's web-protection feature, if one exists. That is worth a note to whoever owns that firewall.
Now change one fact. Suppose the connection log had shown the web server opening a new outbound session to an unfamiliar address on an unusual port. In that scenario, this happened three minutes after the alert, and it was something the server had never done before. Question three now fails, and the decision jumps straight to investigate now. What is that destination? What is the server doing? Has anything on it changed? If it looks like a compromise, escalate — isolate the host, preserve the logs, bring in whoever your incident process names. Same alert, same rule, same priority; the surrounding data made the entire difference, which is the whole lesson of triage.
Where to go next
You do not need a security team to read a sensor. You need five questions. Is the target relevant? Did the connection complete? Did the target talk out afterward? Is the source internal? Has this fired before? Answer them from inventory, flows, DNS, and monitoring history you mostly already have. Every alert ends in dismiss, watch, investigate, or escalate. Every dismissal that will recur becomes a disable, a threshold, or a suppression. And fifteen minutes a day plus an hour a week keeps the whole thing readable. The IDS/IPS for Generalists pillar continues with the maintenance habits that keep a sensor tuned and trusted. The small-team detections guide shows how to turn the patterns you learn here into standing rules of your own.
Put this into practice with SysPeep Monitor
Three of the five triage questions are answered by your monitoring, not your sensor. They are what the target is, what it normally does, and whether it did anything different afterward. SysPeep Monitor keeps that history ready in one web console — discovered devices, per-interface traffic, host health over SNMP and WMI. So the two-minute enrichment look really takes two minutes. The free trial installs on a Windows machine.
Download Free TrialLearn About SysPeep