A zero day vulnerability is a flaw exploited before its fix is public, and the label depends on which dates you can see
Four sources use the term four ways
NIST's entry is about an attack, the others about a vulnerability. NIST's words do not say unknown to whom: the vendor, the public or the defender. A defender cannot see the vendor's knowledge, which is why the working tests below rest on dates.
| Source | Wording | Test it implies |
|---|---|---|
| NIST CSRC glossary, "zero-day attack" (from CNSSI 4009-2022 and NISTIR 8011 Vol. 3) | "An attack that exploits a previously unknown hardware, firmware, or software vulnerability." | Test A: the flaw was unknown when it was first used |
| Google Threat Intelligence Group, 2024 zero-day analysis | "GTIG defines a zero-day as a vulnerability that was maliciously exploited in the wild before a patch was made publicly available." | Test B: first exploitation is earlier than the public fix |
| Microsoft Security Update Guide record | Two fields at release, publiclyDisclosed and exploited, and a release status such as "Exploitation Detected" | Test C: the vendor marks exploitation on the day the fix ships |
| Project Zero, vendor deadlines | Not a definition. For evidence of exploitation against real users "in the wild", details are disclosed after 7 days, against a normal 90 | Evidence of use in the wild shortens the vendor's time to fix |
Three dates decide which test a bug passes
Collect three dates from the advisory and the reports around it: E, the earliest dated evidence of exploitation; V, when the vendor learned of the flaw; F, when a fix became public.
- E earlier than F: a zero day under Test B.
- E earlier than V: the strict sense of Test A, because the vendor could not have known. V is rarely public, so it usually comes from the vendor's own account.
- E after F: not a zero day under either test. It is an n-day, a flaw exploited after its fix is public.
- E on the release day with no earlier date: record it as exploited at release, earlier use not shown, and do not call it either way.
BeyondTrust CVE-2026-102764 shows E, V and F for one bug
BeyondTrust's own timeline gives every date. Counted from E, 29 Aug 2026: the vendor first saw evidence after 30 days, the patch came after 40 days and disclosure after 47 days. The same bug is a zero day under Test A for 30 days and under Test B for 40.
A third reading applies to a self-hosted customer. The patch existed from 8 Oct, but a customer who had not been told had no reason to apply it until 15 Oct. The zero day ends for you on the day you learn there is something to apply.
| Date (2026) | Event | Role |
|---|---|---|
| 29 Aug | The actor exploits the flaw against one BeyondTrust-hosted instance and one self-hosted appliance | E |
| 27 Sep | Another vendor's zero day becomes public, which starts a hunt in BeyondTrust's threat intelligence platform | |
| 28 Sep | The hunt finds the hosted instance; it is contained within 27 minutes of detection; Product Security reproduces the entry route that evening | V |
| 30 Sep | Mandiant shares findings from the self-hosted customer; BeyondTrust validates a fix on a test build | |
| 8 Oct | BT26-05 patch released to SaaS and self-hosted customers; SaaS instances patched automatically | F |
| 15 Oct | Coordinated disclosure of CVE-2026-102764 and the public advisory | Disclosure |
Adobe's Style Smuggler was exploited three days before its hotfix
Sansec reports that attacks on CVE-2026-75650 started on 4 Sep 2026. It published its findings on 5 Sep and wrote that the attackers were iterating quickly. Adobe published APSB26-146 on 7 Sep at 20:20 UTC, with the fix shipped as hotfix VULN-39341 rather than a new release number.
E is 4 Sep and F is 7 Sep, so the bug passes Test B with three days of margin. Unlike the BeyondTrust case, an outside firm watching stores supplied the evidence, and the public record gives no V.
Four more cases, each read against the tests
The record decides the label, and the label should say what the record shows. The PHP row passes Test B on attempts only; the other three do not pass it.
Scroll the table sideways for more columns
| CVE and product | Fix or advisory | Earliest exploitation record | What the record supports |
|---|---|---|---|
| CVE-2026-109618, PHP | 21 Oct 2026, 14:00 UTC | 21 Oct, 03:59 UTC: first request from a named tool; 05:11 UTC: first exploit kit (Fastly telemetry) | Attack tools about ten hours before the fix. Fastly notes that success is not visible at the edge, so the evidence is of attempts |
| CVE-2026-68820, Windows afd.sys | 11 Aug 2026 | 11 Aug: Microsoft marks exploited Yes and publicly disclosed No at release | Exploited at release; earlier use not shown, so Test B cannot be applied |
| CVE-2026-34486, Apache Tomcat | 9 Apr 2026 | 24 Apr, the earliest date in SOCRadar's range of 24 to 29 Apr | An n-day, exploited 15 days after the advisory |
| CVE-2026-9198, Langflow | Release 1.10.1 on 23 Jun 2026; IBM bulletin 2 Jul | CISA's entry in NVD, stamped 17 Jul, lists exploitation as active; catalog listing 4 Aug | An n-day: the fix was public weeks before any evidence of use |
Where the usual answer is wrong
- Unknown to the vendor is not something a runbook can test. Write the trigger in observable terms: a dated report of exploitation, a fix date, and the gap between them.
- A low rating does not rule out exploitation, and a high one does not make a zero day. Broadcom rates ESX CVE-2026-41709 Low (CVSS 2.7) and CISA still added it to its catalog on 28 Oct 2026. The rating describes the flaw; the dates describe the event.
- This site uses the label in observable terms only. The Disclosure Clock on the home page records disclosure, first dated exploitation and the days between, and does not call a bug a zero day; the playbooks start their hour count at the advisory, so for a bug exploited earlier, Question 6 of the 72-hour review sets the window to search.
- A patch on the shelf does not end the exposure for hosts that have not applied it. BeyondTrust's SaaS customers were patched on 8 Oct without action; self-hosted appliances were not.
Related pages
Sources
- NIST CSRC glossary: zero-day attack, read 19 Nov 2026
- Google Threat Intelligence Group: A 2024 Zero-Day Exploitation Analysis, read 19 Nov 2026
- Microsoft Security Update Guide record for CVE-2026-68820, read 19 Nov 2026
- Project Zero: deadlines for vendors, read 19 Nov 2026
- BeyondTrust: ARMORY CASTLE and BT26-05, read 19 Nov 2026
- Mandiant: ARMORY CASTLE Exploits BeyondTrust, read 19 Nov 2026
- Sansec: StyleSmuggler research, read 19 Nov 2026
- Fastly: PHP2Shell timeline and telemetry, read 19 Nov 2026
- SOCRadar: SNOWLIGHT campaign, read 19 Nov 2026
- CISA Known Exploited Vulnerabilities catalog, read 19 Nov 2026
- IBM security bulletin for Langflow OSS, read 19 Nov 2026
Last reviewed 19 Nov 2026.