Reading width
Wide uses the full column for everything, text, diagrams, code, and exercises. Narrow keeps the standard reading width.
Text size
Scales the body text. Headings and code blocks keep their size.
In this section
Incident Findings and Breach Notification Deadlines: GDPR, SEC and NIS2
Introduction
An investigation produces query results. What the organization needs is findings: statements about what happened, precise enough to act on and honest enough to defend. Between the two sits a translation that is easy to get wrong in both directions, by claiming more than the records support or by hedging so heavily that nobody can act.
The translation matters because findings travel. They go into a notification to a regulator, a statement to customers, a briefing to the board, an insurance claim and, sometimes, legal proceedings. Each of those readers will rely on the words, and some will test them against the evidence months later.
This lesson covers what a finding is and the four parts it needs, how to state confidence, how to state an absence, the notification clocks that most incidents start, and how one set of facts is written for several different readers. The deadlines here are summaries of the law for orientation, and an organization's lawyers decide how they apply.
From a Result to a Finding
What the records say, and what you can sayA query result is a statement about records: these rows exist in this table for this window. A finding is a statement about the world: this person's mailbox was read by someone who was not them. The gap between those two is where investigators earn their keep, and it is bridged by reasoning that the finding should make visible.
The course dataset gives a clean example. The question is what the attacker did in the compromised finance mailbox, and one query answers the first part of it:
OfficeActivity
| where UserId == "c.richardson@ne.com" and Operation == "MailItemsAccessed"
| where ClientIP startswith "185.234.72."
| summarize Records = count(), Items = dcount(OfficeObjectId),
First = min(TimeGenerated), Last = max(TimeGenerated) by ClientIP
Two rows. From 185.234.72.18, 20 records covering 20 distinct items between 27 February and 3 March. From 185.234.72.41, 28 records covering 28 items, all in the early hours of 2 March. That is the result. The finding built from it says more, and says how it got there:
Finding 3: mailbox access through the stolen session
Statement
Between 27 February and 3 March 2026, the attacker read at least 48 distinct items in c.richardson's mailbox through the stolen session.
Basis
48 MailItemsAccessed records, each naming one item, from two addresses tied to the attacker's sign-ins.
Confidence
High: direct records of item access, from addresses established as the attacker's by independent sign-in evidence.
Limits
Covers itemized access only: folder synchronization and throttled periods are not itemized. Excludes access by the application the attacker consented to, which is Finding 4.
The words "at least" are part of the finding, not a hedge. The records establish 48 items and cannot rule out more, because folder synchronization and throttled periods are not itemized, as lesson 0.2 described. Saying "48 items were read" would be a stronger claim than the evidence supports, and someone checking it would be right to challenge it.
Notice also the step the finding takes that the query does not. The query shows item access from two addresses; the finding attributes that access to the attacker. The attribution rests on a separate line of evidence, the sign-in records that established those addresses as the attacker's, and the finding says so in its basis. Every finding contains steps like this, from records to a claim about people, and the ones that go wrong are the ones where the step was taken silently.
The Four Parts of a Finding
Statement, basis, confidence, limitsEvery finding an investigation relies on should have the four parts the example shows. Leaving any of them out is how findings fail later.
The four parts of a finding, in the order they are written. A finding missing any of them fails when somebody asks a second question.
One claim per finding is the rule that makes the rest work. "The attacker accessed the mailbox and exfiltrated payment data" is two claims, with different evidence and probably different confidence, and joined into one sentence they take the confidence of the weaker. Splitting them forces each to stand on its own basis.
A first draft of the mailbox finding shows how quickly the structure is lost:
Finding 3, as first drafted
FailsWhat is wrong with it
- Two claims in one: access and theft need separate evidence
- "Stole" is not what MailItemsAccessed records; it records items opened or synced
- No basis, no confidence and no limits: the first question breaks it
The version in section one fixes all three by splitting the claim, naming the record and stating what it does not cover.
Specific wording matters as much as structure. "Accessed" and "read" are different claims, since a synchronized folder was accessed without anyone necessarily reading its contents. "Exfiltrated" claims data left the organization, which needs evidence of an outbound transfer, not only access. "Compromised" is vague enough to mean anything. Choosing the verb the records actually support, and using it consistently, is the cheapest way to make a finding defensible.
The basis must be findable. "Per the audit logs" is not a basis. "48 MailItemsAccessed records in OfficeActivity, query q-021 in the case folder" is. The collected evidence from lesson 0.4 is what the basis points to, and a finding whose basis cannot be traced to a collected export is the first one to fall when challenged.
The limits are where honesty lives. Every finding has them, and naming them protects the finding rather than weakening it. A regulator reading "we cannot determine which items were synchronized, because synchronization is recorded per folder" knows exactly what the organization knows. One reading "48 items were accessed", who later learns about synchronization, starts doubting everything else in the report.
Stating Confidence
The highest level the evidence will carryAbsolute certainty is rare in a tenant investigation, and findings that pretend otherwise are fragile. The aim is to state each finding at the highest level of confidence the evidence will carry when somebody tests it: not the level the investigator believes, and not the most cautious level that could be defended, but the level the records support.
What moves a finding up is corroboration: independent records that agree. The mailbox finding above is high because two independent kinds of evidence point the same way. Sign-in records establish the Singapore addresses as the attacker's, and separate audit records show item access from exactly those addresses. A finding resting on one record, or reached only by ruling everything else out, sits lower.
The corroboration behind the mailbox finding can be shown in one query:
let u = "c.richardson@ne.com";
let net = "185.234.72.";
union withsource = Source
(SigninLogs | where UserPrincipalName == u and IPAddress startswith net),
(AADNonInteractiveUserSignInLogs
| where UserPrincipalName == u and IPAddress startswith net),
(OfficeActivity | where UserId == u and ClientIP startswith net)
| summarize Records = count(), First = min(TimeGenerated) by Source
Three sources, written by different services, place the same account at the same addresses from the same evening. The sign-in logs establish whose addresses they are; the mailbox records establish what was done from them. Neither depends on the other, which is what makes the finding high.
Confidence climbs with corroboration. A finding moves up a step only when a record that does not depend on the others agrees with it.
Module 1 sets out the four-level calibration this course uses and the environment-specific version of it, and the scale should be printed in every report, because "high" and "medium" are ordinary words that readers will otherwise define for themselves. The habit to build now is assigning a level to every sentence in a draft finding. Sentences that resist a level usually turn out to contain two claims, or a claim with no basis at all.
Some readers will push for certainty that the evidence cannot give. An executive who asks "so was data stolen or not" deserves a direct answer, and a direct answer can still carry its confidence: "almost certainly some mail was read; we cannot yet say whether any of it was copied elsewhere." That sentence is more useful to a decision than a yes or a no that might not survive the next day's evidence.
Confidence is also allowed to change. An early finding stated at a middle level can move up when more evidence is collected, or down when a record turns out to mean something else. Recording each change, with the evidence that caused it, is part of the decision log from lesson 0.1, and it is what lets a report explain why an early statement to a regulator was later revised.
Elimination deserves a particular warning, because it feels stronger than it is. "The attacker must have entered through the phishing link, because no other route appears in the logs" is a finding reached by ruling things out, and it is only as good as the logs' coverage. If the non-interactive sign-ins were not routed, or the device the user clicked from was not monitored, the other routes were never visible to rule out. A finding reached by elimination states what was eliminated and from which sources, so a reader can see the size of what was not.
Stating an Absence
"No evidence of" needs its conditionsThe most consequential findings are often negatives. Was data sent outside the organization? Did the attacker reach other mailboxes? Were any payments changed? A clean "no" can decide whether a notification is needed, and it is the claim most often made carelessly.
The course dataset shows how to make one properly. The question is whether the attacker sent mail from the compromised mailbox through the stolen session:
// Returns 0: no Send records from the attacker's addresses.
OfficeActivity
| where UserId == "c.richardson@ne.com" and Operation == "Send"
| where ClientIP startswith "185.234.72."
| count
The count is zero. Before that becomes a finding, three conditions from earlier lessons have to be checked. Was the record type produced? Yes: Send is audited in this tenant, and the same mailbox has 77 Send records over the month, all from its owner. Was the window retained? Yes: the records reach back to 13 February, before the attacker arrived. Could the query find it? Yes: run without the address filter, it returns the owner's own sent mail, so the query works.
With all three established, the finding can be stated, with its conditions:
Finding 6: no mail sent through the stolen session
Statement
No mail was sent from c.richardson's mailbox from the attacker's addresses between 27 February and 5 March 2026.
Conditions
Send is audited for this mailbox; records cover the full window; the query returns the owner's sent mail when the address filter is removed.
Limits
Does not cover mail copied out by the forwarding set on 3 March, which uses no Send record. That is Finding 7.
The limits line is what makes this finding safe. The forwarding rule sent copies of every incoming message outside the organization from 3 March, and that left no Send record because the transport service did it, not the mailbox. A report stating "no mail left the mailbox" would have been false. The narrower, conditioned statement is true, and it points the reader to where the outflow actually was.
The general principle has a well-worn phrase: absence of evidence is not evidence of absence. In a tenant it has a precise meaning. An absence becomes evidence only when the source would have recorded the event if it had happened, the records cover the period, and the search was capable of finding it. Those are exactly the three conditions above. When all three hold, a negative finding can be stated with high confidence. When any one fails, the honest statement is that the question cannot be answered from the available records, and why.
This distinction often decides notification. "We found no evidence that data was copied" and "we cannot determine whether data was copied, because file synchronization is not itemized" lead to different decisions about whether people must be told, and only the second is true in many tenants.
The Clocks
Deadlines that start at awarenessMany incidents start one or more legal or contractual clocks, and most of those clocks start running at awareness, a point that comes much earlier than certainty. The investigator does not decide whether a deadline applies; legal, privacy and compliance colleagues do. But the investigator controls the evidence those decisions rest on, and needs to know which facts start which clocks.
The common clocks, measured in hours from the point each one starts. Most start at awareness; the SEC clock starts at the materiality decision, which itself must be made without unreasonable delay.
The meaning of awareness is the part that most affects investigators. European data protection guidance treats a controller as aware once it has a reasonable degree of certainty that a security incident has compromised personal data. It expects the initial investigation to establish that quickly, and it does not allow the clock to wait for the full investigation to finish.
The clocks are also not all the same kind. The GDPR clock to the regulator has a risk threshold: a breach that is unlikely to result in a risk to people need not be notified, though it must still be documented internally. The communication to individuals has a higher threshold of high risk. The SEC clock starts at a materiality determination rather than at discovery, but the rule requires that determination to be made without unreasonable delay, so it cannot be postponed to postpone the filing. United States state breach laws, health privacy rules and sector regulators add their own definitions and periods, some measured in days and some in weeks.
What they share is that each turns on facts the investigation produces: whether personal data was involved, how many people, what kind of data, whether it was accessed or only exposed. The investigator's job is to establish those facts quickly and state them precisely, so that the people deciding about notification are deciding on findings rather than impressions.
In the course dataset's main incident, that point arrives early. Once the investigator has read the Singapore sign-in and the mail access that followed it, there is a reasonable degree of certainty that an outsider read a finance mailbox, which will contain personal data. The 72 hours can be running before anyone has worked out how the attacker got in.
Once the time of awareness is recorded, the deadlines are arithmetic, and worth computing rather than estimating. For an illustrative awareness time of 09:40 UTC on 5 March:
print Aware = datetime(2026-03-05T09:40:00Z)
| extend NIS2EarlyWarning = Aware + 24h,
GDPRAndNIS2Notification = Aware + 72h
The early warning falls due the next morning and the 72-hour notifications on Sunday 8 March, a weekend, which is when a deadline worked out in someone's head is most often missed.
That is why lesson 0.1 insisted on declaring early, and on recording the time. The decision log's first entries, when the incident was declared and on what evidence, are often the evidence of when awareness began. Writing them accurately, in UTC, matters more than it seems at the time.
Phased Statements
Saying what is known, and when more is comingThe clocks above rarely allow a complete account. Regulations anticipate that: GDPR allows information to be provided in phases where it is not all available at once, and most regimes expect an initial notification followed by updates. The skill is writing an initial statement that will not need to be retracted.
An initial statement built from findings, rather than from impressions, does that naturally. It states what is established at high confidence, names what is still being investigated, and commits only to what the evidence supports. The mailbox finding in section one, stated as "at least 48 items", can be updated upward if the synchronization records show more. A statement that said "a small number of emails" would have to be withdrawn.
Each update follows the same discipline. It restates what was established before, says what has changed and why, and keeps the earlier wording wherever it still holds. A sequence of statements that each correct the last is damaging, but a sequence that each add to the last, with numbers that only ever move in the direction the "at least" anticipated, reads as an investigation working properly.
Initial statement, as sent
PhasedWording that sounds reassuring is a particular trap. "There is no evidence that data was exfiltrated" sounds like a finding and is often only a statement about where nobody has yet looked. Every negative in an external statement should be one of the conditioned absences from section four, or it should not be there.
One Set of Facts, Several Readers
Different emphasis, never different factsThe same incident is described to very different readers: the board, a regulator, affected customers, the insurer, the IT team that has to change the tenant. Each needs different emphasis and different detail. None of them may be given different facts.
Four readers, four documents, one source. The emphasis changes from reader to reader; the facts never do.
The practical way to keep the facts consistent is to write the findings once, in the four-part form, and build every audience's document from them. A number that appears in the board summary, the regulator notification and the customer letter should be the same number, traced to the same finding, with the same "at least" attached. Inconsistency between documents is one of the first things a regulator, a journalist or an opposing lawyer notices, and it damages trust in all of them.
One finding, built into two documents:
Finding 3, written for two readers
For the board
An attacker read at least 48 emails in one finance mailbox over five days. We are confirming whether more were copied and will update on Friday.
For the regulator
Unauthorized access to one mailbox, 27 February to 3 March 2026; at least 48 messages accessed; categories of data subjects under assessment.
The dates, the count and the "at least" survive both translations. Only the vocabulary and the emphasis change.
The documents differ in vocabulary as well as length. A regulator's form asks for categories of data subjects and records, so the finding about 48 messages becomes a statement about the categories of personal data those messages contain, which means someone has to read them or sample them. Executives need the business consequence: whether payments were affected and whether customers must be told. The technical team needs operation names, addresses and object identifiers. Translating the same finding into each vocabulary, without changing its substance, is a skill Module 9 practices directly.
Legal privilege is a further consideration in many organizations: some investigation documents may be prepared at the direction of counsel to keep them privileged. That decision belongs to the lawyers, and it affects what the investigator writes down and where. Module 9 covers it, and the habit to start with is asking early whether counsel is directing the investigation.
What Module 9 Builds On It
Reports, notifications and closureModule 9, Post-Incident and Communication, builds the full discipline. It covers the report that outlives the evidence, where the causal chain actually stops, what the collection lets you notify, three clocks in one incident, the numbers somebody will check, privilege over an export, four readers and one set of facts, the metric that improves as evidence gets worse, turning findings into tenant changes, and what "closed" actually means.
The habit to take into it is to write findings as you go rather than at the end. A short query that counts what an investigation has established per account, kept up to date, becomes the backbone of the findings list:
OfficeActivity
| where ClientIP startswith "185.234.72."
| summarize Records = count(), First = min(TimeGenerated), Last = max(TimeGenerated)
by UserId, Operation
| order by First asc
Each row is a candidate finding: one account, one kind of action, a count and a window. On the course dataset it returns four rows for the finance account: the inbox rule and its update on 27 February, 48 mail item accesses running to 3 March, and the forwarding change on 3 March. Each needs its statement, basis, confidence and limits written before the report is drafted, and each operation that does not appear, such as Send, needs its conditioned absence.
Practice
Would you sign this statement? prepared evidenceThe prepared evidence below is a draft regulator notification for a mailbox compromise, with the investigation's findings beside it. Find each sentence the evidence does not support and say how it should be written.
The second path is to write findings yourself from the course dataset, in the four-part form, including at least one conditioned absence.
- Run the checklist query in section eight and pick two operations to turn into findings.
- Write each in four parts, statement, basis, confidence and limits, using "at least" wherever the records cannot rule out more.
- Write one absence finding following section four, checking all three conditions before stating it.
- Write a two-sentence executive summary built only from those findings, with no new claims.
A finding is one claim about what happened, with its basis, its confidence and its limits. Confidence rises with independent corroboration, absences need their conditions checked and stated, and notification clocks usually start at a reasonable degree of certainty, long before the investigation is complete. Every audience gets the same facts from the same findings, with only the emphasis changed.
The last lesson in this module turns to recovery: what Microsoft 365 can restore, for how long, and what it cannot.
Resources
Laws and guidance are summarized here for orientation; an organization's lawyers decide how they apply.