Insider risk · 3.6

False positives, and what they cost

False positives, and what they cost. What the requirement says, what it means in practice, and what an assessor will ask.

For an independent reference point, see NIST SP 800-53 Rev. 5.

The cost

What a wrong alert actually costs

The obvious cost is time: somebody investigates something that turns out to be a project deadline.

The larger cost is credibility. A queue that is mostly noise gets processed less carefully each week, and the mechanism by which a real signal is missed is not carelessness but rational adaptation to a system that has been wrong two hundred times.

The third cost is the worst. An alert acted on too quickly, involving a person who did nothing wrong, damages that person and teaches everybody watching what the programme is for.

Where they come from

Four sources

Baselines built too quickly. A fortnight of observation does not capture month-end, quarter-end or the annual audit.

Rules written for a different organisation. Default detections assume an office. A manufacturer moves large files to machines at odd hours as a matter of routine.

Roles not distinguished. What is unusual for an accounts clerk is Tuesday for a programmer.

Legitimate change. A new customer, a new production line, a new tool. The system correctly reports that things are different.

Tuning

An activity, not a setting

Observe before alerting, for long enough to see the cycles. Suppress known patterns explicitly and with a reason recorded. Write rules per role rather than globally. And review the alerts that were closed as nothing, monthly, looking for the pattern that produces them.

That review is the work. Systems that are never tuned converge on being ignored, which is the same as not having them.

What to measure

Precision, not volume

The useful number is the proportion of alerts that turned out to be worth investigating. A system producing four alerts a month of which three merit a look is more valuable than one producing four hundred of which five do.

Vendors report volume because it demonstrates activity. Ask instead what proportion of alerts their customers act on.

Handling one properly

Quietly first

Check the obvious explanations before speaking to anybody: is there a project, a deadline, a shift change, a system migration. Ask the person's manager whether the activity fits the work.

Only then, if it still does not fit, does anybody talk to the person, and that conversation is a question rather than an accusation.

This sequence costs half an hour and prevents almost every case where these programmes cause harm.

Recording the outcome

Including the ones that were nothing

What was alerted, what was checked, what the explanation was, who decided. The record demonstrates a functioning process to an assessor and it is the input to the monthly tuning review.

It also protects the person who was investigated and cleared, which matters if the question ever arises again.

The target

Few alerts, each worth reading

An organisation of a hundred people should expect a small number of signals a month, not a dashboard. If the system produces more, it is untuned rather than the organisation being unusually eventful.

Saying so is against the interest of anybody selling by the alert, and it is what makes the capability worth having.

Starting quiet

Deploy in observation mode first

Collect without alerting for a full business cycle, then look at what would have fired. Almost always the answer is far too much, and tuning against real data before anybody is asked to act on an alert is the cheapest possible version of this work.

Also

Elsewhere in insider risk