How many CVE records will NIST's National Vulnerability Database list as published in calendar year 2027?

TL;DR

The saved forecast is 10th percentile: 129,750 CVE records calendar year 2027. Treat this as a lower-tail threshold, not an expected total or a hard floor. The decisive evidence is the sustained rise in late-2026 publications and the separation between NVD listing and NIST analysis.

Context

The question counts distinct, non-rejected CVE records whose NVD publication timestamps fall within calendar year 2027. The reference dataset is the one available on January 15, 2028. Identifier year, last-modified date, severity scoring, and completion of NIST analysis do not determine eligibility. The measurement uses publication-date windows and excludes Rejected records through the official interface (NVD API documentation).

The forecast date is November 5, 2026. The evidence shows a publication system operating at a much higher volume than in earlier years. At the same time, discovery, validation, CVE assignment, upstream publication, NVD ingestion, and enrichment remain different stages. More activity at one stage does not guarantee an equal increase at the next.

Evidence

The historical backbone is the NVD publication-date series. Non-rejected annual totals were 25,073 in 2022, 28,816 in 2023, 39,953 in 2024, and 48,150 in 2025. January–October 2026 already contained 89,852 eligible records. The full monthly history below shows both the older baseline and the change during 2026. Each cell is all statuses / non-rejected; the observations are historical counts, not forecasts (official NVD API).

Publication month	2022	2023	2024	2025	2026
January	2,207 / 2,016	2,565 / 2,337	2,647 / 2,594	4,415 / 4,275	5,143 / 4,302
February	2,017 / 1,942	2,474 / 2,122	2,906 / 2,778	3,831 / 3,673	4,808 / 4,616
March	2,092 / 2,058	2,783 / 2,516	3,352 / 3,309	4,166 / 3,998	6,304 / 6,234
April	2,063 / 2,046	2,363 / 2,330	3,735 / 3,621	4,142 / 4,034	5,885 / 5,810
May	2,035 / 2,021	2,734 / 2,418	5,103 / 4,999	4,264 / 3,980	6,985 / 6,936
June	2,331 / 2,279	2,405 / 2,384	3,173 / 3,078	3,799 / 3,681	8,001 / 7,942
July	2,156 / 1,969	2,363 / 2,305	3,174 / 3,113	3,933 / 3,776	9,919 / 9,770
August	2,409 / 2,317	2,518 / 2,477	2,945 / 2,896	3,737 / 3,628	12,719 / 12,256
September	2,323 / 2,197	2,828 / 2,148	2,540 / 2,514	4,537 / 4,321	15,074 / 14,928
October	1,909 / 1,849	2,722 / 2,680	3,595 / 3,564	4,389 / 4,273	17,481 / 17,058
November	1,989 / 1,958	2,508 / 2,449	4,086 / 4,056	3,116 / 3,020	Incomplete
December	2,900 / 2,421	2,686 / 2,650	3,448 / 3,431	5,643 / 5,491	Not observed
Total; January–October for 2026	26,431 / 25,073	30,949 / 28,816	40,704 / 39,953	49,972 / 48,150	92,319 / 89,852

The table shows why a forecast anchored mainly to the previous annual total would miss the current state of play. Early 2026 was higher but still recognizable against the older monthly range. The later months were different. Non-rejected publications rose each month from May through October, reaching 17,058 in October compared with 4,273 in October 2025 (official NVD API). I read this as evidence of a higher publication baseline, not proof that the latest growth rate will continue indefinitely.

Partial November observations do not establish a reversal. One available window contains 1,712 non-rejected records through November 4; another, longer window contains 1,829 through November 5 in its available snapshot. Their coverage differs, and neither is a completed month (official NVD API). The completed-month sequence carries more weight than a few days of releases.

The next decisive fact is institutional. NIST’s April 15, 2026 policy keeps submitted CVEs flowing into the database while selecting records for immediate enrichment. NIST also reported a 263% increase in submissions between 2020 and 2025 and nearly one-third year-over-year growth in first-quarter 2026 submissions. Those are submission statistics, not interchangeable publication counts (NIST announcement). An August 12 notice describes automated ingestion within approximately an hour of CVE publication, followed separately by analyst enrichment (Federal Register notice). Together, these facts argue against imposing an analyst-capacity ceiling on this question.

Upstream demand supports persistence, but not unlimited conversion. GitHub describes rising vulnerability-report and advisory workloads alongside processing delays. The evidence supports pressure on the pipeline; it does not supply a global report-to-NVD conversion rate (GitHub operational report). Linux’s recurring process for assigning CVEs to potentially security-relevant fixes in supported stable trees also supports continuing issuance rather than an explanation based entirely on a one-time historical backlog (Linux CVE process).

AI-assisted discovery adds another source of pressure. Anthropic’s November 2 dashboard reported 37,966 candidates, 6,376 externally reviewed findings, 6,009 confirmed findings, 5,087 disclosures, 629 patches, and 230 CVE records. These are vendor-reported workflow counts, not a representative global conversion study (Anthropic dashboard). Their main lesson is the gap between finding something and publishing a CVE. Human validation, duplicates, maintainer decisions, and coordinated disclosure all intervene.

There are concrete counterweights. GitHub introduced reporting limits and structured forms on October 1, 2026. These controls target intake quality; they are not a universal CVE publication cap (GitHub reporting controls). Cisco describes bundling multiple fixes into CVEs associated with weakness categories, showing that more discovered bugs need not produce the same increase in record counts (Cisco announcement). The CVE Program’s Quality Era framework likewise emphasizes governance, infrastructure, and record quality rather than announcing a numerical quota (CVE Program announcement).

Other mechanisms point in both directions. Updated CNA rules address prompt publication and rejection, while the frontier-AI researcher CNA pilot tests a new issuance pathway under validation requirements. Neither establishes unrestricted automated publication or a measured future throughput gain (CNA rules, pilot description). Research into reserved-but-public identifiers confirms delayed-publication stock exists, but incomplete coverage prevents treating it as a fixed addition to the target year (reserved-but-public research).

What's non-obvious

The relevant bottleneck is not necessarily NIST analysis. A record can count while awaiting analysis, deferred, or otherwise unenriched. The obvious story—an overwhelmed NVD must publish fewer eligible records—confuses listing with enrichment. The policy and ingestion evidence separate those functions. That distinction is central to explaining why the lower-tail threshold remains elevated despite visible analysis constraints.

The opposite shortcut is also wrong: more AI findings do not mean proportionally more CVEs. Validation, deduplication, bundling, disclosure timing, and assignment rules determine how discoveries become records. This question measures administrative records, not independently verified underlying vulnerabilities. The information advantage lies in tracking that conversion chain while giving recent completed publication months more weight than older annual averages.

Uncertainties

The largest uncertainty is persistence. The late-2026 rise is observed; the share attributable to recurring inflow rather than backlog clearance is not established. Queue-age distributions, remaining backlog inventories, and publication volumes by major assigning organization would help distinguish those explanations. Broad post-control intake and conversion data would also show whether stricter reporting requirements are reducing noise or slowing valid disclosures.

Rejection revisions are a separate risk. October 2026 contained 17,481 records before filtering and 17,058 after filtering. January’s difference was much larger: 5,143 versus 4,302 (official NVD API). A single fixed rejection percentage would miss that variation. Records eligible today can be rejected before the reference measurement, and timestamp corrections or later additions can move historical totals in either direction.

Finally, the available evidence does not establish the adoption rate of new publication automation, the eventual scale of AI-CNA participation, or a complete global discovery-to-publication conversion rate. Those gaps support uncertainty, not numerical additions. The saved 10th-percentile forecast remains a lower-tail assessment that allows substantial retreat from the recent surge without treating that retreat as the main reading of the evidence.