Read a sample: 1. What Detection Engineers Do
On a Tuesday morning at the fictional company Northstar Design, an analyst notices an alert: an employee account had several unsuccessful sign-in attempts, followed shortly by a successful sign-in.
That sequence deserves attention. Someone may have been trying passwords until one worked. But there are other explanations. The employee may have mistyped a password, remembered it, and signed in successfully. A phone might have been using an old saved password while the employee signed in from a laptop. The records alone do not tell the analyst what the person behind the activity intended.
The alert exists because someone decided that this pattern was worth bringing to an analyst’s attention. They determined which records could describe the behavior, wrote logic to find a version of it, and considered how to check whether that logic worked. That work is detection engineering.
Detection engineering turns a concern about security into a way to look for observable evidence. It is not simply the act of writing a query, and it is not the same as proving that an attack occurred. A useful detection is a testable approach to finding activity that warrants attention, along with an understanding of the data it needs and the situations in which it may mislead.
This chapter introduces the people, records, and decisions involved. You do not need to know a particular security platform or programming language. The goal is to build a clear mental model of how suspicious behavior becomes something a team can monitor.
Organizations generate records as people and systems go about their work. An employee signs in to an application. A server starts a scheduled job. A laptop runs a program. A user changes a setting. A service rejects a request.
A security event is a recorded occurrence that may be useful for understanding security-relevant activity. It can describe something ordinary, unsuccessful, administrative, or suspicious. The word security does not mean the event itself is malicious.
For example, Northstar’s identity system might produce these fictional records:
- At 9:02 a.m., a sign-in attempt for the account morgan.lee failed.
- At 9:03 a.m., another sign-in attempt for morgan.lee failed.
- At 9:05 a.m., a sign-in attempt for morgan.lee succeeded.
Each record describes an occurrence. Together, they suggest a sequence. Neither the individual records nor the sequence establishes who controlled the account or why the failures happened.
Events are evidence, but evidence has limits. A record can be incomplete, delayed, incorrectly configured, or difficult to interpret without other records. Even a well-formed event usually captures only the perspective of the system that produced it. An authentication service may know that a password was accepted, for instance, without knowing whether the person who entered it was the legitimate account owner.
A detection engineer starts with those limitations in mind. The question is not merely, “Can we find failed sign-ins?” It is, “What pattern involving sign-ins would be useful to surface, what evidence would we need, and what might cause us to misunderstand it?”
A detection is a defined method for identifying activity of interest in available data. It includes a description of what the team is looking for and logic that can be applied to relevant records. In practice, teams may also use the word to refer to a rule, the surrounding documentation, or the broader monitoring capability. The important idea is that a detection connects a security concern to observable conditions.
At Northstar, the concern might be:
> Someone could gain access to an employee account after making repeated unsuccessful sign-in attempts.
That is a reasonable concern, but it is not yet rule logic. The account-access claim involves intent and identity: who was trying to sign in, and whether they were authorized. The authentication records may not answer either question directly.
The detection engineer can instead look for a more limited, observable pattern:
> Several failed sign-ins for one account are followed by a successful sign-in for that account within a short period.
This description does not claim the activity is an attack. It identifies a pattern that could be consistent with the concern and could also have benign explanations.
To turn that description into a working detection, the engineer must make further choices. What counts as “several”? How short is the period? Must the attempts come from the same source, or should activity from different sources be connected? Which sign-in events are included? What happens when an account name is missing?
Those choices affect what the detection finds and what it misses. There is no universally correct threshold that can be selected without considering the available data and the environment. A detection engineer makes a starting choice, tests it, examines the results, and revises it when the evidence supports a change.
An alert is a notification or item created to draw attention to activity. A detection may generate an alert when its conditions are met, although teams can also use detections in other ways, such as reviewing matches in a report or using them to enrich an investigation.
Suppose Northstar creates a rule with this fictional starting condition: find three or more failed sign-ins for an account, followed by a successful sign-in for the same account within ten minutes of the first failure. When that condition is met, the monitoring system creates an alert.
The alert is not another word for the underlying activity. It is the result of applying a decision rule to recorded activity.…