Log for the questions you will ask
The chapter opens with a Sunday alert: a Meridian administrator account signed in from Lagos four minutes after signing in from a site in the Midlands. The alert let the team disable the account before the intruder acted, and the logs let them reconstruct what it had touched. Visibility depends on logging, but noise costs attention, the scarcest resource a small team has.
The book suggests choosing log sources by working backwards from the questions an investigation will ask: who authenticated, from where and with what result; which privileged actions were taken; what data was accessed or moved in bulk; what changed in the cloud and identity configuration; and what the security tools saw. That gives a rough priority order: identity (sign-ins including failures, MFA events, privileged-role activations, directory changes), the cloud control plane (Azure Activity and AWS CloudTrail), endpoint and EDR events, email security, network sources such as firewall, DNS and VPN or ZTNA authentication, and finally critical applications and databases.
Deciding what not to log matters as much. Secrets, full card numbers, passwords (a failed attempt can contain a real password typed into the wrong field), session tokens and unnecessary personal data turn the log store into a target and into a shadow copy of sensitive data with its own privacy obligations. The book recommends masking sensitive fields at the source, recording the metadata of an event rather than its content wherever that is enough, and separating verbose debug logs from the security audit trail, with different retention for each.
Centralise and protect the record
Logs left on each server can be erased by an attacker on that machine, and nobody will log in to forty systems to build a timeline. Shipping logs as they happen to a central store that the sending system cannot alter solves both problems, and it enables correlation: a sign-in from an unusual place, then a privileged-role grant, then access to a sensitive share, tell a story none of the three systems could tell alone. The chapter shows an rsyslog configuration that forwards over TLS with a disk-backed queue, so a brief collector outage does not drop events, and names Windows Event Forwarding and the cloud diagnostic-settings pipelines as the equivalents elsewhere.
Treat the central store as a high-value asset: limit who can query it, limit far more tightly who can delete from it, and apply a retention lock so a stolen admin credential cannot shorten how long evidence survives. Retention should cover realistic attacker dwell time, months rather than thirty days. To control cost, keep recent weeks hot and searchable and move older logs to cheaper immutable archive, giving identity and control-plane logs the longest retention and writing the policy down. Synchronise clocks and store timestamps in UTC.
Running a SIEM without drowning
A SIEM ingests the central logs, supports search and correlation, and runs detection rules. Microsoft Sentinel is the usual choice for Microsoft 365-centred organisations, with the Elastic stack and Wazuh as open-source options. The hard part is noise: teams switch on many built-in rules, drown in false positives and learn to ignore the channel.
The suggested discipline is to start with a few high-confidence detections, treat every alert as having a cost, tune or remove rules that routinely fire on benign activity, and read an unmanageable alert volume as a tuning problem rather than a staffing problem. Detections are placed in three tiers: a very small set that pages a person immediately, a larger set that feeds a calm daily or weekly review queue, and everything else kept as searchable history. For more experienced readers, the book recommends building detections on attacker behaviour rather than indicators such as IP addresses or hashes, in the spirit of MITRE ATT&CK, and keeping detection rules in version control with saved sample events so they can be tested after changes.
Three starter detections and the problem of baselines
The chapter provides read-only KQL for three early detections: impossible travel (successful sign-ins by the same user from different countries less than two hours apart), a new or newly elevated administrator (directory audit events adding a user to a privileged role), and mass download (a user whose file downloads today exceed both a floor of 200 and five times their 14-day daily average). Each describes a behaviour, compares against a baseline or an impossibility, and ends with a human decision rather than an automatic action.
Baselines adapt to the environment but drift, so thresholds need periodic review and generous floors, and some false positives are the accepted price of catching attacks no signature anticipated. At Meridian the first weeks of Sentinel produced forty to fifty alerts a day; an afternoon spent adding exceptions for known VPN egress ranges, replacing a noisy service-account match with a narrower rule, and tightening impossible travel brought this down to three or four real items a day. That is why the Lagos alert landed in a channel the team still read.
From dashboards to action
Dashboards are what people ask for first and matter least. The one the book considers worth building is a coverage-and-health view showing whether every expected source is still reporting, because a source that goes silent is itself a finding. It extends this to meta-detections on the monitoring pipeline: a normally hourly source that has sent nothing for several hours, ingestion volume for a key source falling by an order of magnitude, or a change to a diagnostic or audit-logging setting.
The author describes a maturity path from collection, to a few trustworthy detections that are acted on, to proactive hunting, and advises against leaping to hunting before the second stage is solid.
KEY POINTS
Key points
- List the questions you would need to answer after an incident and confirm the identity, cloud, endpoint, email and network logs that answer them are collected.
- Forward logs over TLS to a central store the sending systems cannot alter, with a disk buffer so short outages do not drop events.
- Put a retention lock on the log archive and set retention in months, keeping identity and control-plane logs longest.
- Enable impossible travel, new or elevated admin, and mass download detections, and tune them to your own baselines.
- Review alert volume and cut or tune rules until every alert is one someone would act on.
- Alert when an expected log source goes silent, and write a short runbook note for each detection.
QUESTIONS
Common questions
- Which logs should a small IT team collect first for security?
- Start with identity logs: sign-ins including failures, MFA events, privileged-role activations and directory changes. Next come the cloud control plane (Azure Activity, AWS CloudTrail), endpoint and EDR events, email security events, and network sources such as firewall, DNS and VPN authentication, then critical applications and databases. The Hardened Stack recommends choosing sources by the questions an investigation will ask rather than collecting everything.
- What should not be written to security logs?
- Avoid logging secrets, full payment-card numbers, passwords, session tokens and more personal data than you need, because the log store then becomes a target and a copy of sensitive data with its own privacy obligations. Even failed sign-in attempts can contain a real password typed into the wrong field. Where possible, mask sensitive fields at the source and record who, when, where and the outcome rather than the content of the event.
- How long should security logs be retained?
- Long enough to cover realistic attacker dwell time, which means months rather than thirty days, because many intrusions are found long after initial access. A practical approach is to keep recent weeks hot and searchable and move older logs to cheaper immutable archive with a retention lock. Identity and cloud control-plane logs usually justify long retention; verbose application logs rarely do.
- How do you stop a SIEM from producing too many alerts?
- Start with a small number of high-confidence detections instead of enabling every built-in rule, and tune or disable any rule that routinely fires on harmless activity. Separate detections that page someone immediately from those that feed a daily or weekly review queue, and keep everything else as searchable history. If alerts exceed what the team can review, treat it as a tuning problem rather than a staffing problem.
- What are good first detections to build in Microsoft Sentinel?
- The Hardened Stack suggests three: impossible travel (the same account signing in from different countries within a short window), a new or newly elevated administrator, and mass download compared against each user's own recent baseline. All three look for attacker behaviour rather than specific indicators, and each should end with a person reviewing the result. Thresholds such as minimum download counts need tuning to the environment.
- How can I tell if a log source has stopped sending logs?
- Build a coverage-and-health view that shows whether every expected source is still reporting, and alert when a source goes silent. Add checks for a sharp drop in ingestion volume and for changes to diagnostic or audit-logging settings. An attacker who disables logging relies on nobody noticing the silence, so a silent source should be treated as a finding.