Summary

  • Seven yearly rows in DIVD's "Some statistics" table sum exactly to the 200 cases and the 1.433.802 addresses on its homepage. The arithmetic closes; the definitions do not, because no reviewed page says what a case or a vulnerable IP is.
  • The CVE Numbering Authority page prints its intervals beneath a sentence that twice reserves the right to deviate, and the three case pages reviewed on 21 November 2026 print a status and a modification time but no first-publication date.

Start with a case that closed DIVD-2026-00002. The metadata block reads Status "Closed", Last modified "18 Nov 2026 20:17 CET", case lead Victor Pasman, patch status "Available", the two identifiers CVE-2026-1281 and CVE-2026-1340 against Ivanti Endpoint Manager Mobile. The dated timeline runs: 29 January 2026, initial publication of the vulnerabilities; 4 February 2026, the advisory updated to include public proof-of-concept availability; 5 February 2026, the case started by DIVD researchers, instances scanned, notifications sent; 7 August 2026, case closed. The page's own chart records the case span as 5 February to 7 August 2026, 183 days.

That is a better case record than most regulators publish. You can reconstruct the shape of the work from it. Two things sit outside it, and they are the subject here: the earliest timeline entry, 29 January, precedes the chart's start date by seven days, so "how long did this case run" depends on which printed start you choose; and the only temporal field beside the status is a modification time. When the page went public is not printed on it.

The page that prints the case

Across the three case pages reviewed on 21 November 2026 — DIVD-2024-00019, DIVD-2026-00002 and DIVD-2026-00014 — the pattern holds. The Operation Endgame page prints "Status Closed", "Last modified 07 Aug 2026 12:30 CEST", CVE(s) "n/a", a likely-abuse window of 1 December 2023 to 30 May 2024, 30 May 2024 as the day the Dutch National Police went public and DIVD sent its first notifications, and 7 August 2026 as the close; its chart gives a span of 799 days. The open Zammad case prints "Status Open", "Last modified 18 Nov 2026 20:09 CET", first unauthorised access on 21 September 2026, DIVD awareness on 22 September 2026, a case-file update on 18 November 2026, and the two identifiers CVE-2026-102489 and CVE-2026-102490. Each gives you a status, a timeline and a modification stamp. None of the three prints a field saying when the case file first appeared.

That is a bounded finding and it is worth stating narrowly: on the three pages reviewed, not across the whole case index, "Last modified" is the only clock in the metadata block. A reader who needs to know when a stranger could first read the file has to guess from the modification time, which only moves forward.

The page that prints the clock

Now the policy. The CVE Numbering Authority page prints the intervals anyone would want: triage "Within 2 weeks after submission", vendor notification "Within 1 week after successful triage", notification treated as received after 30 days of sending the first notice where effort was sufficient and no acknowledgement came, the case file published 30 days after the initial disclosure where acknowledgement fails, a product warning at 60 days after acknowledgement for software and 90 days for hardware or firmware, and full disclosure of details "typically postpose" — the page's own spelling — to at least 6 to 12 months after the vendor released a patch.

Above all of it sits the sentence that governs all of it, and it appears twice on the page: "Please note that the timelines given below are an indication and not set in stone. We reserve the right to deviate from these timelines if this better suits our mission." The Code of Conduct says the same thing in a different register — disclose to the vendor first, then request CVE numbers and "negotiate a reasonable time span" for telling partners and the public — and adds that earlier disclosure is possible where a vendor is obviously slow and abuse is likely.

Read together, these are two descriptions of one practice: a default schedule with an explicit override, and a negotiated interval with no day count. Neither is a deadline a stranger can hold anyone to, and neither claims to be. The interesting part is not the discretion; disclosure norms generally carry it. The interesting part is that the schedule is printed to the day while the record that would show whether it ran is not. A reader can count 30 days from a notification date on a case timeline. Nothing on the case page says which of the printed intervals the case was actually run against.

The page that prints the count

The DIVD CSIRT homepage carries a table headed only "Some statistics", five columns — Year, # of cases, # of vulnerable IPs notified, # of leaked credentials notified, # of cves assigned/requested — and seven rows. 2020: 14 cases, 58,358 addresses, 386 leaked credentials, 0 CVEs. 2021: 25, 99,006, 5,000,100, 24. 2022: 42, 297,472, 0, 19. 2023: 37, 375,531, 0, 15. 2024: 45, 509,098, 1,218,000,000, 28. 2025: 28, 90,304, 156,843,612, 32. 2026: 9, 4,033, 27,642,432, 27.

Here is the arithmetic I ran myself, and it is the only place in this piece where DIVD's numbers are added. The seven case figures total 200. The seven address figures total 1,433,802. DIVD's homepage prints, in one verbatim line, "189 Members 200 Total cases 1.433.802 Vulnerable IPS Notified". The front page and the table agree to the unit.

That agreement is a real thing to check, and it passes. It is also where a stranger's arithmetic has to stop, for four reasons that are all printed absences. No as-of date accompanies the counters. No cut-off date accompanies the 2026 row, which prints 9 cases on a page that also lists more current case references than nine. No sampling frame or scanned-population denominator appears beside the table, so the address column cannot become a rate.

And no counting rule defines a case or a vulnerable IP: the leaked-credentials column, which swings from 386 to zero to 5,000,100 to 1,218,000,000 to 156,843,612, makes that absence vivid, since nobody reading the page can tell whether a credential is an account, a record or a line in a dump.

Three surfaces, one year

The institute does not only count this way. The extracted text of the DIVD Annual Report 2024 states the caseload grew to 52 in 2024 from 37 in 2023, that 508,391 IP addresses were notified, that two notable cases involved informing 26 million victims of stolen credentials, and that 28 CVEs were assigned and published. The extracted text of the 2025 financial report states 18 CSIRT cases in 2025 leading to 46,405 notifications on vulnerable IP addresses, compares that with 52 cases and 508,391 addresses in 2024, states stolen-credential data processed of 1,568,436,120 in 2025 against 26 million in 2024, and records 70 matters handled by the CNA, 47 of them in the energy track.

Set those beside the table and the gaps are exact. For 2024 the table prints 45 cases and 509,098 addresses; the annual report's extracted text prints 52 and 508,391. Seven cases and 707 addresses. For 2025 the table prints 28 cases and 90,304 addresses; the financial report's extracted text prints 18 and 46,405. Ten cases and 43,899 addresses. On leaked credentials, the 2025 table figure of 156,843,612 and the financial report's 1,568,436,120 differ by exactly a factor of ten.

I am not going to tell you that is a decimal point: no reviewed document says one figure restates the other, and the nouns differ — credentials notified in one, data processed in the other. The factor is a fact about two printed numbers, not a diagnosis.

The honest conclusion is duller and more useful than a scandal. Three surfaces of one institution answer the same question — how much did we find in 2024 and 2025 — and they do not agree, because each is counting something the publisher has not defined in the text I read. A tally that closes internally and three tallies that do not line up are not a contradiction. They are what a record looks like when the rule lives in someone's head.

Where my own arithmetic fails

One comparison I tried and could not make, published here as a failure rather than quietly dropped. The newsroom article "Over 1 Million Notifications" states that DIVD notified vulnerable organisations of at least 1,295,494 compromised IP addresses across 143 unique cases. The homepage prints 200 cases and 1,433,802 addresses. Two case counts, two address counts, one institution. I cannot order them, subtract them or reconcile them, because the inspected excerpt prints no date and no measurement window for the article's figures. 143 is not a subset of 200 that I can demonstrate, and I will not pretend otherwise. The missing field defeats the comparison before the arithmetic starts.

Who decides, and who is named

The Responsible Disclosure page is worth one line to keep it out of the way: it governs reports about DIVD's own systems, promises a response within three business days with an evaluation and an expected resolution date, and sets no outbound publication interval. It is not the disclosure policy; the CNA page is.

On the CNA page, case files are maintained by "The DIVD CNA Admins". In the text reviewed, no individual is named as the person who decides to publish a given case. Divisions of labour inside a volunteer institute are not a scandal either — a role designation is a normal way to publish. But it means the chain between a printed interval and a specific decision has no named end. What the case pages do name is the operator: Victor Pasman as case lead on DIVD-2026-00002, Frank Breedijk on DIVD-2024-00019 according to that page's metadata, the DIVD Crisis Management Team on DIVD-2026-00014. The person doing the work is credited.

The person deciding to disclose is a plural.

What would make this checkable

Four cheap fields, and none of them requires anyone to disclose more about a vulnerability than they already do. A counting rule printed beneath the statistics table, one sentence per column. An as-of date on the counters and a cut-off on the current year's row. A first-publication timestamp on the case page, separate from "Last modified", so the interval a reader measures is the interval the policy describes. And, beside each case, the name of the interval it was run under — 30 days, 60, 90, 6 to 12 months — so the discretion clause becomes a recorded choice rather than an untestable permission.

BTW's directory entry for Stichting Dutch Institute for Vulnerability Disclosure records a Dutch foundation with no prior BTW coverage. My own adjacent work read statutory and standards clocks — the GDPR breach articles, the Cyber Resilience Act's Article 14, RFC 7258, the CA/Browser Forum incident guidelines — where the instrument tells you what to count and who must be told. This is the other direction: an institute that publishes more per case than most regulated firms, and leaves the counting rule to the reader.

Nothing above alleges that a disclosure was late, that a vendor was treated unfairly, that a count is inflated or that a case was mishandled. The record reviewed on 21 November 2026 shows the opposite of opacity for individual cases: names, dates, identifiers, patch status, an open case updated in public while it is still open. What it does not show is the rule. In a disclosure ecosystem that increasingly borrows other people's aggregates, the rule is the part that travels.