Leaving this course?

Lesson 1 of 12

What the function is for

A security operations function has three jobs, and most of its published advice concerns only the second. It has to be able to see what is happening, decide whether what it sees matters, and act in time for the decision to be worth anything.

Why framing it as a product fails. The tooling market sells the middle job — correlate, alert, dashboard. But an alerting platform fed by the wrong data cannot see, and a platform whose alerts nobody can action does not decide. Organisations routinely buy the middle and discover the function still does not work, because the first and third jobs were never anyone's responsibility.

The published frameworks place this correctly. The current cybersecurity framework, version 2.0, describes six concurrent and continuous functions: Govern, Identify, Protect, Detect, Respond and Recover. Detection is one of six, and it sits between protection, which decides what is worth watching, and response, which is what detection exists to trigger. A detection capability with nothing downstream is an expensive way to generate records of things you did not prevent.

The question that tells you whether the function works. Not "how many alerts did we handle". It is: for the last incident, or the last credible near-miss, what told us, how long after the activity started, and who decided it mattered? If the answer is that an outside party told you, or that the alert existed but nobody triaged it, the function has a specific fixable fault rather than a resourcing shortage.

What this course does not assume. It does not assume a dedicated team, a 24-hour rota or a budget. Most of what determines outcomes is the collection decisions in lessons three to five and the detection discipline in lessons six and seven, all of which can be done by one competent person with access to the right data. The organisational material in lesson twelve matters when you are scaling, not when you are starting.

What it does assume. That you would rather know the uncomfortable numbers than report comfortable ones. Several lessons here end with a measurement that is usually worse than the organisation believes — how far back the logs actually go, how many alerts are closed without being read, how long the function would take to answer a specific question. Those are the numbers that change anything.

The shape of what follows. One documented case of a detection that worked, examined closely, because it is more instructive than any framework. Then collection, retention and telemetry integrity. Then detection engineering, which is the core skill. Then triage, hunting, measurement, the limits of visibility, and the sustainability problem that the available evidence says is real.

Lesson 2 of 12

The detection that worked

In June 2023 one organisation detected a nation-state actor using a forged authentication token, and it was the only one of twenty-two affected organisations to do so. A statutory board documented exactly how, which makes this the most useful worked example available.

The sequence. On 15 June 2023 the organisation's security operations centre detected anomalies in access to its mail systems. The next day it saw multiple alerts from a custom rule it had written and named internally "Big Yellow Taxi". The rule analysed a log recording access to mailboxes. The team investigated, determined the activity was malicious, triaged it as a moderate-level event, and contacted the provider on 16 June.

Four things had to be true, and all four are decisions rather than purchases.

First, they held the right log. The board recorded that the log in question "was not accessible without that 'premium' service", which the organisation had purchased. Collection is lesson three.

Second, they had written their own rule against it. The board heard that few organisations analysed that log in detail and that such analysis "would be difficult for smaller organizations". The log was necessary; the rule was what worked.

Third, the rule was built from experience rather than from a template. The board recorded that the team "designed custom alerting capabilities based on three years of experience dealing with anomalous access to mailboxes." Specifically, it curated the log data "to enumerate all applications accessing mailboxes within its infrastructure, and to trigger alerts for any anomalous events", and designed a rule "to detect deviations in mailbox activity by comparing baseline interactions of applications with Exchange Online."

Fourth, somebody investigated an ambiguous alert. This is the part most easily lost. The board noted that the alerts "showed activity that could have been considered normal", and that the organisation "had seen false positive Big Yellow Taxi detections in the past". A team under volume pressure closes that alert. This one did not.

What the rule actually did, in plain terms. It enumerated which applications were touching mailboxes, established what normal looked like for each, and alerted on deviation. That is not a signature for a known attack — no signature for this attack existed, because the technique was a forged token from a stolen signing key. It was a model of normal, which is why it caught something nobody had seen before.

The lesson to carry into the rest of this course. The detection that worked was not bought, was not a signature, and was not obvious when it fired. It was the product of holding specific data, understanding one narrow behaviour deeply, and having an analyst willing to investigate something that looked like it might be nothing. Every subsequent lesson is about making those three things reliably available.

Lesson 3 of 12

Deciding what to collect

Collection is where the function is won or lost, and it is the decision most often made by cost rather than by value. This lesson is how to make it deliberately.

The definition worth starting from. The current draft of the national log management guidance defines a log as "a record of events that occur within an organization's computing assets, including physical and virtual platforms, networks, services, and cloud environments", and frames the work as a lifecycle of "generating, transmitting, storing, accessing, and disposing of log data." Five stages, each of which can fail independently — and the one that usually fails silently is transmitting.

A published model for prioritising sources. A 2021 federal memorandum establishes a maturity model across four event logging tiers: EL0, rated not effective, where the highest-criticality logging requirements are not met; EL1, basic, where only the highest-criticality requirements are met; EL2, intermediate, adding the intermediate requirements; and EL3, advanced, where requirements at all criticality levels are met.

The useful part is how criticality is defined. "Each log category has an assigned criticality level based on its relative cybersecurity value. This cybersecurity value relates to the usefulness of the log data for threat detection, with the most useful data assigned a criticality of zero, and the least a criticality of 3." That is the right organising idea: rank log sources by their usefulness for detection, not by their volume, their availability or whether a product happens to collect them.

What the tiers actually require, which makes the model usable rather than abstract. The basic tier covers logging categories, minimum logging data, a time standard, event forwarding, protecting and validating log information, passive DNS, access arrangements for investigating authorities, and planning for both orchestration and user behaviour monitoring. The intermediate tier adds further logging categories, publication of a standardised log structure, and inspection of encrypted data.

Note what appears at the basic tier. A time standard, and protecting and validating log information. Both are unglamorous and both are prerequisites: correlating events across systems whose clocks disagree produces false conclusions, and telemetry you cannot show to be intact is evidence you cannot rely on. Lesson five is the second of those.

What to do with this if you are not a federal agency. Use the structure, not the mandate. List your log sources. Assign each a criticality from zero to three on the single criterion of usefulness for detecting an attack. Then establish, honestly, which criticality-zero sources you are not collecting at all. In most organisations that list is short, specific and immediately actionable — and it is a far better spending case than a request for a bigger platform.

The source most often missing. Identity. The documented case in lesson two turned on a log of access to a service, and the identity provider is the system whose compromise matters most. If your criticality-zero list does not begin there, it is ranked by habit rather than by value.

Lesson 4 of 12

Retention, and the number that decides investigations

How far back your logs go determines which investigations are possible. Most organisations do not know the real figure, and the real figure is usually much shorter than the policy says.

A published benchmark. The same federal memorandum sets, for logs of the highest criticality, 12 months in active storage followed by 18 months in cold data storage — thirty months in total. It defines active storage as data "stored in a manner that facilitates frequent use and ease of access", and cold data storage as storage "in a manner that minimizes costs while still allowing some level of access and use."

Compare that against what you have. Thirty days of searchable logs is a common commercial default and a frequent budget outcome. Against a thirty-month benchmark it is roughly three per cent of the window. The gap matters because intrusions are routinely discovered months after they begin, and because the documented case in lesson two illustrates the consequence precisely: when the provider enabled enhanced logging for victims who had not purchased it, they could see forward from that moment and no earlier. The intrusion window stayed dark.

Why the two-tier structure is the right answer rather than a compromise. Searchable storage is expensive and investigations need depth, not speed, after the first week. Splitting retention lets you afford a long tail: recent data fast and queryable for detection and triage, older data cheaper and slower for investigation and for answering the question "how long has this been happening?" An organisation that cannot answer that question cannot scope an incident, and cannot tell a regulator or a customer what happened.

The deadlines the memorandum set, which are worth knowing as evidence that this is achievable rather than aspirational: the basic tier within one year of the memorandum, the intermediate tier within eighteen months, and the advanced tier within two years. A government mandated a three-year retention posture on a two-year timetable.

The three questions to answer this week.

  • For each criticality-zero source, how many days of data can you actually query right now? Not the policy figure — run the query.
  • How many days exist in cheaper storage that you could retrieve if required, and has anyone ever retrieved any?
  • If you were asked today how long a given account had been behaving unusually, what is the earliest date you could speak to?

The common and expensive failure. Retention configured per source, by whoever deployed each one, with no overall figure. The function then discovers mid-incident that the one source it needs rolled over eleven days ago. Produce a single table of source against actual retention, and you will find at least one surprise. That table is also the cheapest thing in this course to make.

Lesson 5 of 12

Protecting your own telemetry

Your logs are evidence, and an attacker who can alter or stop them has removed your ability to establish what happened. This lesson is the small set of controls that make telemetry trustworthy, and they are among the least implemented in this course.

The requirement that almost nobody has. The federal memorandum states it directly: "When logging stops unexpectedly, audit alerts shall be sent in near real-time to any parties responsible for monitoring. The responsible party must promptly investigate the cause of the disruption and take appropriate corrective actions."

Ask whether you would know. For most organisations a log source that stops reporting simply becomes quiet, and quiet is indistinguishable from peaceful. The absence of events is itself an event, and it is both the most common operational failure in this field and a recognised attacker behaviour. A detection for "source X has sent nothing for N minutes" is cheap, applies to every source, and is the single highest-value rule most functions are missing.

Integrity, as the memorandum frames it. It requires "conducting integrity checks periodically and upon access against the log hashes throughout their retention period", and treats logs that "are no longer being written to or are considered closed" as a distinct state. Hashing at rest and verifying on access turns a log store into something you can assert about, which matters when the question is whether an attacker edited the record rather than whether they triggered an alert.

Why this is not paranoia. The whole value of the function rests on the record. If an intrusion involved administrative access — and the documented cases in this course mostly did — then the logs were reachable by the attacker's credentials. A central store that the compromised identity could write to or delete from is a record whose completeness you cannot vouch for, and you will discover that during the investigation rather than before it.

The practical controls, in order of value.

  • Alert on silence, per source, with a threshold tuned to each source's normal cadence.
  • Forward off the originating host promptly, so that destroying a machine does not destroy its history.
  • Write to a store the ordinary administrative credentials cannot delete from — append-only where available, and separately administered where not.
  • Hash and verify, so that integrity is a check rather than an assumption.
  • Monitor for configuration changes to logging itself, because disabling collection is the step before the activity you wanted to see.

The one that is free. Alerting on silence requires no new product and no additional data. It uses telemetry you already pay to collect and tells you when you have stopped collecting it. If you adopt one thing from this course, adopt that.

A note on monitoring configuration change. The memorandum also requires "monitoring across the enterprise for unexpected changes to files or configuration items". Applied to the logging pipeline itself, this closes the loop: an attacker who turns off a source should generate both a silence alert and a change alert, from two independent paths. Controls that depend on the thing they are watching are not controls.

Lesson 6 of 12

Detection engineering

Detection engineering is the discipline of building, testing and maintaining detections as a product rather than configuring a tool. It is the core skill of this function and the one least often treated as a skill.

Why a public catalogue of adversary behaviour is the right backbone. The widely used framework organises real observed techniques into tactics — the attacker's objective at each stage. Its current enterprise version lists fifteen: Reconnaissance, Resource Development, Initial Access, Execution, Persistence, Privilege Escalation, Stealth, Defense Impairment, Credential Access, Discovery, Lateral Movement, Collection, Command and Control, Exfiltration, and Impact.

Note two of those names. In the current version the tactic numbered TA0005 is called Stealth, and Defense Impairment (TA0112) is a separate tactic of its own. Material written against earlier versions calls the former Defense Evasion and has no separate tactic for the latter. If your coverage map, your reporting or your training uses the older names, it is describing a superseded version — and the split is substantive, because degrading a defence and hiding from it are different behaviours needing different detections.

How to use the catalogue without deceiving yourself. As a coverage map, listing which techniques relevant to your estate you could detect, with evidence. The failure mode is counting techniques: a map claiming coverage of three hundred techniques, where coverage means a rule exists, tells you nothing about whether any rule fires on real activity. One technique with a tested detection beats fifty with untested ones.

The complementary catalogue. A countermeasure knowledge graph, published by the same organisation and funded by a national security agency, encodes defensive techniques and "supports queries that can inferentially map cybersecurity countermeasures to offensive tactics, techniques, and procedures." Its value is the direction of travel: from the control you have to the behaviour it addresses, rather than from the attack to a control you might buy.

The cycle, which is what makes this engineering. Form a hypothesis about an observable behaviour. Establish what data would show it and confirm you hold that data. Write the rule. Test it by generating the behaviour deliberately and confirming the rule fires. Tune it against real traffic until the false positive rate is survivable. Document what it detects, what it does not, and what an analyst should do when it fires. Then re-test on a schedule, because the estate changes underneath it.

The step that is usually skipped. Testing by generating the behaviour. An untested detection is a hypothesis with a name, and the most common discovery on first testing is that the rule never fires at all — a field was renamed, the log source does not carry the attribute assumed, or the activity never reaches the collector. Writing the rule is perhaps a fifth of the work.

Why the documented case in lesson two is detection engineering rather than configuration. The team enumerated which applications touched mailboxes, baselined each, and alerted on deviation. That is a hypothesis about an observable, implemented against a specific data source, informed by three years of knowing what normal looked like. No product ships that. It was built.

Lesson 7 of 12

Writing a rule that survives

Most detections fail in production rather than in testing, and they fail in a small number of recognisable ways. This lesson is how to write one that lasts.

The two failure modes. A rule that fires constantly is switched off, formally or informally, and a rule that never fires is indistinguishable from one that does not work. Both end as entries in a coverage map describing a capability you do not have. Everything below is about staying between them.

Detect behaviour, not artefacts. A rule keyed to a filename, a hash or a specific address stops working the moment the attacker changes it, which costs them nothing. A rule keyed to a behaviour — a service account authenticating from a new location, an application reading mailboxes it has never touched, a volume of outbound data outside the established range — keeps working because the behaviour is what the attacker needs rather than what they chose. The documented case in lesson two is precisely this: deviation from a baseline of application behaviour, not a signature.

Baselines, and the thing nobody says about them. A baseline needs enough history to be meaningful and needs maintaining as the estate changes. It is also a privacy and proportionality question when the behaviour is a person's rather than a system's, which is a reason to prefer service and application baselines where either would work. They are more stable, there are fewer of them, and they raise fewer questions you would have to answer to a regulator.

Write for the analyst who will receive it at three in the morning. Every rule should carry, in the alert itself: what behaviour triggered it, why that behaviour is suspicious, what would make it benign, the first three things to check, and who to escalate to. A rule without this is a rule whose value depends on the author being awake. This is the cheapest quality improvement available to a detection function and it is skipped almost universally.

On false positives, honestly. The documented case recorded that the organisation "had seen false positive Big Yellow Taxi detections in the past" and that the alerts, when the real intrusion occurred, "showed activity that could have been considered normal". A rule with no false positives is usually a rule that is too narrow to catch anything novel. The goal is not zero; it is a rate low enough that each alert is still investigated, and a known benign pattern that is documented rather than rediscovered each time.

Tuning is a record, not an act. Every exclusion should say why it exists, who approved it and when it should be reviewed, because an undocumented exclusion is how a rule quietly stops covering the thing it was written for. The exclusion list is the part of a detection most likely to be exploited and least likely to be audited.

The retirement question. Rules should be removed when the behaviour no longer applies or the data source is gone. A detection library nobody prunes becomes a library nobody trusts, and the cost is not storage — it is that the function cannot say which of its detections are real.

Lesson 8 of 12

Triage and the economics of alerts

A function fails in triage more often than in detection. This lesson is the arithmetic, because the problem is usually structural rather than a matter of effort.

Do the sum for your own team. Take your daily alert volume, multiply by the honest average minutes to reach a confident disposition, and divide by the analyst hours actually available for triage — excluding meetings, projects and the other work the same people do. If the result exceeds the hours you have, the surplus is not being triaged. It is being closed, deferred or ignored, and which of those is happening is a question worth answering explicitly rather than discovering during an incident.

Why adding analysts rarely fixes it. Volume scales with the estate and with every new detection; headcount does not. A function whose answer to volume is always more people has an economics problem that will recur every year. The durable fixes reduce volume per unit of coverage rather than increase capacity.

The four things that actually reduce volume.

  • Raise fidelity at the source. A rule producing fifty alerts a day of which one is interesting should be rewritten, not staffed. Rewriting it is cheaper than a year of triage.
  • Aggregate related alerts into one case. Thirty alerts from one host sequence is one investigation, and tooling that presents it as thirty is generating work rather than information.
  • Automate the retrieval, not the decision. Enrichment — who owns this host, is this address known, what else did this account do in the last hour — is mechanical, and doing it automatically before a human looks is the largest honest saving available. Deciding whether the activity is malicious is judgement, and automating that is how a function stops noticing novel things.
  • Delete detections that have never been right. This is lesson seven's retirement question, and it is the fastest reduction nobody authorises.

The signal that tells you triage is broken. The proportion of alerts closed without any investigative action recorded. If it is high, the function is not triaging; it is clearing a queue. This is the same reasoning as a human reviewer who never overrides an automated decision — a reviewer who never finds anything is not exercising judgement, and the control exists only on paper.

The structural point about tiering. Splitting analysts into a tier that closes alerts and a tier that investigates them is common and has a known failure: the first tier is measured on throughput, which is exactly the measure that produces the closure behaviour above. If you tier, measure the first tier on the quality of its escalations rather than on the volume of its closures, and have the second tier sample what the first closed. Without that sampling, nobody in the organisation knows what is being discarded.

What the documented case depended on. An analyst investigating an alert that "could have been considered normal", from a rule with a history of false positives, during a normal working week. No amount of tooling produces that. Capacity and incentives do.

Lesson 9 of 12

Threat hunting, honestly

Threat hunting is the most oversold activity in this field and also a genuinely valuable one. The difference is whether it produces detections.

What it is. Looking for activity your detections would not have caught, by forming a hypothesis and testing it against your data. "An attacker with valid credentials is using a service account outside its normal pattern — would I see that, and is it happening?" The output is either a finding or the knowledge that you could not have seen it, and the second is as valuable as the first.

What it is not. Browsing dashboards. Running a vendor's content pack. Watching a threat intelligence feed and searching for the addresses in it, which is matching, not hunting — useful, cheap, and not the same activity. Nor is it a substitute for the collection and detection work in lessons three to seven: hunting in data you do not hold finds nothing, and the hunt will not tell you that your retention is thirty days.

The precondition most organisations have not met. Hunting requires data depth and query capability. If you can search thirty days, every hypothesis about slow, patient activity is unanswerable, and the hunt silently restricts itself to what the window permits. That is why retention is lesson four and hunting is lesson nine.

The output that makes it worth the time. A hunt that finds nothing should still produce one of three things: a new detection, so that next time a rule catches it; a collection gap recorded, so the data exists next time; or a documented statement that this behaviour is covered and was checked. A hunting programme producing none of these is an activity, not a capability, however skilled the people are.

Why the "found nothing" result is routinely mishandled. It gets reported as reassurance. It is not reassurance unless you can say what you would have seen, which is why the three outputs above matter. "We hunted for lateral movement and found none" means little; "we hunted for lateral movement, confirmed we hold the authentication data to see it, found none, and wrote two rules so it alerts in future" means something.

A realistic cadence for a small function. One hypothesis a month, chosen from the behaviours most relevant to your estate and documented in half a page: the hypothesis, the data used, the query, the result, and which of the three outputs it produced. Twelve of those a year is a real programme and is achievable alongside other work. A quarterly week-long exercise that produces a slide deck is not.

Where to get hypotheses. From the adversary behaviour catalogue in lesson six, filtered to techniques that apply to your platforms; from the documented cases in this and the other courses, because published incident reviews describe behaviours precisely; and from your own incidents, because the thing that happened once is the thing most likely to happen again and least likely to have a detection.

The honest caveat about scale. Hunting is most valuable where detection coverage is already decent, because otherwise the hypotheses all come back "we would not have seen that" and the right response is to go and build detections. If you are early, spend the time on lessons three to seven. Hunting rewards a function that has already done the unglamorous work.

Lesson 10 of 12

Measuring the function

Security operations metrics are mostly theatre: they rise with activity and fall with nothing. This lesson is the short list that moves only when something real changes.

The measurements worth reporting.

  • Detection provenance for the last incident or near-miss. What told you, how long after the activity began, and who decided it mattered. One sentence, and it is the most honest statement a function can make about itself.
  • Actual queryable retention, per criticality-zero source. From running the query, not from the policy.
  • Silence coverage. What proportion of your log sources would raise an alert if they stopped reporting. Usually near zero before anyone looks.
  • Tested detections. Not how many rules exist — how many have been fired deliberately in a test and confirmed to alert, and when that was last done.
  • Alerts closed with no investigative action recorded, as a proportion. Lesson eight's number.
  • Time to answer a specific question. Pick one, such as "has this account authenticated from an unusual location in the last six months?", and measure how long the function takes to answer it definitively.
  • Hunt output, as the count of hunts that produced a new detection or a recorded collection gap, rather than hunts performed.

Why not mean time to detect and mean time to respond. Because both are computed only over incidents you detected, which excludes the ones that matter most. A function that misses a six-month intrusion and catches a hundred commodity alerts quickly will report excellent figures. The denominator is unknowable, so the average is unsound — and the first measurement above is the honest replacement.

Why not alert volumes or case counts. They measure how much the estate generated and how much the team processed. Both go up when you add a noisy rule, which is the opposite of improvement. Reporting them trains the organisation to value activity.

On maturity models. The current framework describes implementation tiers named Partial, Risk Informed, Repeatable and Adaptive, and the logging model in lesson three has its own four tiers. Both are useful for stating position to a board and for showing a direction. Neither tells you which of your own blind spots is most dangerous this month, which is what the measurements above are for. Use the tiers to report and the numbers to decide.

The uncomfortable framing that works with executives. Do not open with what the function did. Open with what it could not currently see, and what it would cost to change that. A single sentence — "if an administrator's credentials were used from an unusual location four months ago, we could not today tell you" — moves budget in a way that no dashboard does, because it describes a specific, plausible and unacceptable state of affairs.

The discipline to keep. Every number above is one the function might prefer not to produce. That is the test of whether a metric is real: comfortable metrics are the ones that survive without anything improving.

Lesson 11 of 12

Where visibility ends

Every function has a boundary beyond which it cannot see, and most have not drawn it. This lesson is about knowing where yours is, because an unmapped boundary is mistaken for coverage.

The four boundaries.

Software as a service. Your data sits in applications whose logs the provider controls, often at a tier you have not bought. The documented case in lesson two is the clearest illustration: the log that caught a nation-state actor was behind a premium licence, and most victims did not hold it. For each significant service, establish what it logs, what tier you hold, how long it retains, and how you get the data out.

Infrastructure you rent. Covered fully in the cloud course, but the operational point belongs here: the control plane is logged by the provider, those logs are a product decision rather than a given, and switching them on later does not recover the past.

Third parties and managed providers. A supplier with remote access into your estate is part of your attack surface and usually outside your telemetry. The review of the Lapsus$ attacks found that organisations "did not always consider third parties and BPOs in their risk" programmes while attackers specifically targeted those servicers. If a provider administers your systems, their activity should appear in your logs, and whether it does is a question with a yes or no answer.

The provider's own compromise. You cannot detect an intrusion into your provider's infrastructure. What you can detect is its effects in your own data — anomalous authentication, unexpected application access, data movement — which is why lesson six's emphasis on behavioural detection matters more in a rented estate than in an owned one.

The honest limit on knowing how bad things are generally. Two statutory reviews made the same observation from different directions. On Log4j: "no authoritative source exists to understand exploitation trends across geographies, industries, or ecosystems. Many organizations do not even collect information on specific Log4j exploitation, and reporting is still largely voluntary." A European threat landscape report analysed 4,875 incidents in the twelve months from July 2024 — a count of reported incidents, not of incidents. Industry-wide numbers are weaker than they appear, so build your programme on your own exposure rather than on sector statistics.

The exercise that draws the boundary. List your ten most important systems and services. For each, write one line: what telemetry do we receive, from where, how far back, and what would we not see? The lines that read "nothing" or "unknown" are your boundary, and they are almost always a small, specific and purchasable list rather than a structural limitation.

The judgement to apply. A boundary is acceptable if it is known, recorded and accepted by someone with the authority to accept it. It is unacceptable when the organisation believes it is covered. The function's obligation is not to see everything; it is to be accurate about what it sees.

Lesson 12 of 12

Making the function sustainable

This last lesson is about people, because the evidence that the usual model does not hold is in the record rather than in opinion.

The documented finding. The statutory review of the Log4j event recorded that the response "contributed to professional 'burnout' among defenders that may, compounded with the generally intense pace of many cybersecurity jobs, have a long-term impact on the availability of cybersecurity talent." That is a board attributing a workforce effect to a single event, in a report about a software vulnerability.

The structural reason it happened. The same review observed that "cybersecurity professionals cannot collectively sustain a model where most of their resources are spent reacting to emerging problems like a steady stream of vulnerabilities." A function whose capacity is fully committed to reaction has no reserve for the event that requires sustained effort, and the next such event is therefore paid for out of its people.

What that implies for how you staff. Not that you need more people, which is usually unavailable. That you need the routine load to sit below capacity, so that a surge is absorbable. The levers are in lesson eight: fidelity at source, aggregation, automated enrichment, and retiring detections that have never been right. Reducing routine load is a resilience measure, not an efficiency one.

The design decisions that matter most.

  • Do not run a twenty-four-hour rota you cannot staff properly. A thinly covered night shift produces the closure behaviour in lesson eight and burns the people doing it. Honest on-call with a documented response time is better than permanent presence that is awake but ineffective.
  • Write detections for the analyst, not the author. Lesson seven's point: the context in the alert is what lets a tired person at an unsociable hour make a sound decision.
  • Give the function project time. Detection engineering, testing and hunting are the work that reduces future load. A team entirely in triage cannot build the things that would reduce triage, which is the trap.
  • Rehearse before the event. The surge capacity you have is the surge capacity you have practised.

What the whole course has argued, in one paragraph. The one documented detection that caught a nation-state actor was not a product, not a signature, and not obvious when it fired. It depended on holding a specific log, on a rule built from three years of understanding one narrow behaviour, and on an analyst with the time and the standing to investigate something that might have been nothing. Everything in these twelve lessons is in service of making those three conditions ordinary rather than exceptional: collect deliberately, keep it long enough, protect it, engineer detections and test them, keep triage inside capacity, hunt to close gaps, measure what is uncomfortable, know where you cannot see, and run the function at a pace its people can sustain.

The one thing to do first. Alert on silence. It costs nothing, applies to every source, uses data you already pay for, and tells you when you have stopped being able to see. Most functions do not have it, and no function should be without it.

Knowledge check

Fifty questions

Answer all fifty questions, then check your answers. You need 45 out of 50 to pass and receive a certificate. If you score less, you will see which answers were right and wrong, and then go through the course again before you retake the check. Your answers, progress and times are kept only in this browser.

Standards and frameworks covered

The published frameworks this course teaches toward. Saolix is not affiliated with, endorsed by or accredited by any of these bodies; naming a standard is a citation, not a claim about its author.

Sources

The official documents this course relies on. Laws and guidance change, so check the current version.

  1. M-21-31: Improving the Federal Government's Investigative and Remediation Capabilities Related to Cybersecurity Incidents · US Office of Management and Budget, Executive Office of the President
  2. NIST SP 800-92 Revision 1 (Initial Public Draft): Cybersecurity Log Management Planning Guide · US National Institute of Standards and Technology
  3. Logging and protective monitoring · UK National Cyber Security Centre
  4. ATT&CK enterprise matrix · The MITRE Corporation
  5. D3FEND: a knowledge graph of cybersecurity countermeasures · The MITRE Corporation, funded by the US National Security Agency
  6. Cybersecurity Framework 2.0 · US National Institute of Standards and Technology
  7. NIST SP 800-137: Information Security Continuous Monitoring for Federal Information Systems and Organizations · US National Institute of Standards and Technology
  8. NIST SP 800-61 Revision 3: Incident Response Recommendations and Considerations for Cybersecurity Risk Management · US National Institute of Standards and Technology
  9. Review of the Summer 2023 Microsoft Exchange Online Intrusion · Cyber Safety Review Board, US Department of Homeland Security
  10. Review of the December 2021 Log4j Event · Cyber Safety Review Board, US Department of Homeland Security
  11. Review of the Attacks Associated with Lapsus$ and Related Threat Groups · Cyber Safety Review Board, US Department of Homeland Security
  12. Known Exploited Vulnerabilities Catalog · US Cybersecurity and Infrastructure Security Agency
  13. ENISA Threat Landscape 2025 · European Union Agency for Cybersecurity
  14. #StopRansomware Guide, version 3.0 · CISA, Multi-State Information Sharing and Analysis Center, National Security Agency and FBI
  15. Cybersecurity Performance Goals 2.0 · US Cybersecurity and Infrastructure Security Agency
  16. Cloud security guidance · UK National Cyber Security Centre
  17. Cloud security principle 14: secure use of the service · UK National Cyber Security Centre
  18. NIST SP 800-190: Application Container Security Guide · US National Institute of Standards and Technology