Behavioural baselining at runtime is a detection method that learns how each process, host, user and service normally behaves while it is running, then flags live activity that departs from that pattern. It can catch zero-day threats because even an unknown exploit has to do something unusual, such as spawning an odd process or opening an unexpected connection.
Below, we explain how runtime baselining works step by step, what zero-day activity looks like, where the approach falls short and how to control false positives.
Why runtime behaviour matters for zero-day detection
Most traditional controls ask: does this file, hash or domain match something we already know is bad? That works for known threats. It does not work for a zero-day exploit, because by definition there is no signature, no patch and often no public indicator yet.
Runtime baselining asks a different question: is this behaviour normal for this thing, right now? An attacker can hide the exploit itself, but it is much harder to hide what happens next. Once code runs, it has to execute commands, touch files, request privileges or talk to the network. Those actions leave a behavioural footprint, and that footprint can stand out against a well-understood baseline.
If you are new to the wider topic, our earlier overview of how behavioural analysis enhances threat detection covers the general idea. This article goes deeper into the runtime side and the zero-day use case specifically.
How behavioural baselining at runtime works, step by step
1. Decide what to baseline
A baseline is always a baseline of something. Subjects include processes, servers and workstations, user and service accounts, and network segments. Useful signals include parent and child processes, command lines, file and registry writes, loaded modules, outbound connections, authentication events and data volumes.
2. Learn what normal looks like
During a learning period, the system records how each subject behaves. A web server might only ever run its own worker processes and talk to a database. A finance user might work business hours and use a small set of applications. Good baselines also group similar subjects into peer groups, so a new server can be compared with others in the same role, and they account for regular cycles such as month-end processing or overnight backups.
3. Watch behaviour as it happens
Once a baseline exists, live telemetry from endpoints, cloud workloads, identity systems and the network is compared against it continuously. This is the “runtime” part: the comparison happens while activity is under way, not days later during an audit.
4. Score deviations and combine weak signals
A single unusual event is rarely conclusive. Mature approaches score each deviation and then combine related deviations into a stronger picture. One rare process might be noise. A rare process, followed by a new outbound connection, followed by a new scheduled task on the same host within minutes, is a much clearer story.
5. Add context before raising an alert
Context decides whether a deviation is worth a person’s time. How sensitive is the asset? Is the account privileged? Does threat intelligence know anything about the destination address? Has a change ticket been raised for this server? Enrichment turns “something odd happened” into “this matters, and here is why”.
6. Respond, then feed the outcome back
Confirmed threats trigger response actions such as isolating a host or disabling an account. Analyst decisions on false positives are fed back so baselines and scoring improve.
What zero-day activity tends to look like at runtime
Zero-day exploits differ widely, but the behaviour that follows successful exploitation often falls into a few familiar categories:
- Unusual process chains: a server application or document reader starting a command shell, a scripting engine or an unfamiliar binary.
- New network behaviour: a host connecting to a destination it has never contacted, or a server that normally only receives connections suddenly initiating them.
- Privilege changes: new local administrator accounts, unexpected group membership changes or a service running with higher rights than usual.
- Persistence: new services, scheduled tasks or start-up entries appearing outside normal change windows.
- Lateral movement: one machine attempting to authenticate to many others, or using remote administration tools it has never used before.
None of these behaviours proves an attack on its own. Each one, measured against a strong baseline, is a reason to look closer.
Illustrative examples
The following scenarios are illustrative only. They describe typical patterns, not real customers or incidents.
Example 1: a web server that starts a shell
An internet-facing web application has run the same worker processes for months. One evening, a worker process starts a command shell, which then downloads a file and makes an outbound connection to an unfamiliar address. No antivirus signature fires because the exploit is new. Against the baseline, however, three rare events have occurred on one host in a short window, so the activity is scored highly and raised for investigation.
Example 2: a document reader that talks to the internet
A user opens what looks like an ordinary document. The reader application, which normally never launches other programs, starts a scripting engine. Combined with a new outbound connection from the same process, it matches the shape of exploitation and warrants a closer look.
Example 3: a quiet server that starts scanning
A file server normally only receives connections. It suddenly begins connecting to dozens of internal machines on administrative ports. This is the kind of change that network anomaly detection is designed to surface, and it can reveal an attacker who has already used an unknown flaw to get a foothold.
Limitations of runtime baselining
Behavioural baselining is powerful, but it is not magic. Be aware of these limits:
- Poisoned baselines: if an attacker is already present during the learning period, their activity may be learned as normal.
- Slow, quiet attackers: activity that changes gradually, or that closely mimics legitimate administration, can stay within normal bounds.
- Legitimate change: new software, upgrades, restructures and seasonal workloads all look unusual at first.
- Data gaps: missing endpoint telemetry, inconsistent logging or unsynchronised clocks weaken both the baseline and the comparison.
- Living-off-the-land techniques: attackers who use built-in administration tools are harder to separate from real administrators.
How to reduce false positives
- Combine signals instead of alerting on each one. Correlating related deviations on the same host or account is the single most effective way to cut noise.
- Use peer groups. Comparing a server with others in the same role helps distinguish genuine anomalies from role-specific behaviour.
- Feed in change information. Planned deployments and maintenance windows explain many “anomalies” before they reach an analyst.
- Weight by asset and privilege. The same deviation matters more on a domain controller than on a test machine.
- Capture analyst feedback. Every dismissed alert should teach the system something.
- Re-baseline deliberately. After major changes, refresh baselines rather than letting old ones generate a wave of alerts.
How it fits with the Essential Eight
The Australian Signals Directorate’s Australian Cyber Security Centre (ASD’s ACSC) Essential Eight focuses mainly on prevention: patching applications and operating systems, application control, restricting administrative privileges, restricting Microsoft Office macros, user application hardening, multi-factor authentication and regular backups.
Behavioural baselining does not replace any of those controls, and it is not one of the eight strategies. It complements them. Application control and hardening reduce what an unknown exploit can do; runtime baselining helps you notice when something gets through anyway. The logs these controls generate, such as blocked execution attempts or privileged account activity, are also valuable inputs to a behavioural baseline.
How CiBRAI approaches runtime behavioural detection
CiBRAI is an Australian-owned, agentic AI cybersecurity operating platform that brings SIEM, SOAR, threat intelligence, GRC, case management and automated response together, with an always-on AI analyst layer working across them.
For behavioural detection, CiBRAI collects signals from endpoints, cloud, identity and network sources, including the EDR, SIEM and identity tools you already run. Its AI layer groups related events, removes duplicates and adds context, then prioritises incidents using risk, asset sensitivity and business impact. Instead of a list of disconnected anomalies, analysts see one incident story with a timeline and a plain-English summary of what happened and why it matters.
Response runs through guided playbooks and containment automation. You decide which low-risk steps can run automatically and which need human approval, and every action is recorded in the audit trail. You can explore the full flow on the CiBRAI platform overview.
Frequently asked questions
Is behavioural baselining the same as anomaly detection?
They are closely related. Baselining is the process of learning normal behaviour; anomaly detection is the act of spotting departures from it. Runtime baselining simply means both happen continuously on live activity.
How long does it take to build a useful baseline?
It depends on the environment and how regular its activity is. Many systems need to observe at least one full business cycle, including weekly and month-end patterns, before the baseline is reliable. Peer groups can help new systems become useful sooner.
Can behavioural baselining stop a zero-day on its own?
No single control can guarantee that. Baselining improves the chance of detecting exploitation early, but it works best alongside hardening, application control, fast patching once a fix exists, and a tested response process.
Does it replace antivirus or EDR?
No. Signature-based and EDR detections remain valuable for known threats. Behavioural baselining adds coverage for activity those tools do not recognise, and it is most effective when both feed the same investigation workflow.
Want to see how runtime behaviour, threat intelligence and response come together in one place? Book a CiBRAI demo.