Topic 17 · Deep Dive

17.1 What Logs, Events, and Alerts Tell Defenders

Three Different Objects, Often Confused

A log is a continuous stream written by a system: the Linux syslog daemon appends facility-and-severity lines such as auth, cron, or kern at levels from emerg down to debug. An event is one timestamped row extracted from that stream: Windows writes structured records keyed by event ID, where 4624 means a successful logon, 4625 a failed logon, 4688 a process creation, 7045 a new service install, and 1102 the audit log was cleared. An alert is not data at all -- it is a decision. Something compared events against a rule or model and decided a human should look now. Mixing these up is the most common beginner mistake: more logging never produces better security unless something turns rows into decisions.

What a Row Actually Carries

Every useful row has five anchors: a timestamp, a source identity (host, user, process, or interface), an object acted on (file, account, port, registry key), an action verb, and an outcome. Missing any one of them makes the row nearly uninvestigatable, which is why defenders care about configuration as much as coverage. An 4688 event that records only the process ID but not the command line cannot tell you whether powershell.exe ran a benign inventory script or a base64-encoded downloader. Clock skew is the quiet killer of this model: if hosts drift by minutes, the ordering of "failed logon then success then new service" collapses, and correlation engines produce nonsense.

From Rows to Signals

Raw sources -- operating system, network gear, applications -- are noisy on their own. A single workstation can emit tens of thousands of rows a day, of which an adversary-relevant handful exist. Correlation is what compresses that funnel: patterns across time (five 4625 events then a 4624 in sixty seconds), across hosts (one account authenticating in two countries ten minutes apart), or against a baseline (a server that never writes to /tmp suddenly doing so). Because correlation is probabilistic, alerts arrive with a false-positive tail; a defender's real inventory is the ratio of confirmed incidents to tickets, not ticket volume.

Retention Decides What You Can Ever Know

Intrusions are routinely discovered weeks after the first event, so retention window is a detection capability. Keep the fields you will pivot on longest -- auth, process, network flow metadata -- and shrink verbose payload logs faster. Preserve a tamper-evident copy before acting, because attackers with privilege routinely clear event logs (that 1102 event) or delete the very syslog file you are reading.

Architecture Diagram

raw sources to alert OS logs syslog / event IDs Network flows, firewall Apps auth, errors Event rows ts, src, obj, verb Correlate Alert 1 of 40k rows
Monitoring is a funnel: many raw rows are normalized into timestamped events, correlated, and only then become an alert worth a human.

Key Takeaways

« Back to Topic 1717.2 »