Is CVE Growth Exponential, and Does Exploitation Follow?

14 min read #security #research

Contents

Abstract#

The number of published CVE records has grown from about 4,600 in 2010 to 48,151 in 2025, and to 88,650 by 29 October 2026. Using the full CVE List together with CISA’s Known Exploited Vulnerabilities (KEV) catalogue and FIRST’s EPSS scores, I look at the shape of that growth, at who publishes the records, and at how many of them matter in practice. Between 2017 and 2025 the series fits an exponential trend of about 16 % a year better than a linear one. The growth, however, comes in steps, each tied to a change in who can issue CVEs. 2026 breaks that trend: monthly volume roughly quadrupled between January and October. Vendor CNAs, GitHub, VulnCheck and the Linux kernel CNA drive the increase, which coincides with the arrival of AI-assisted vulnerability discovery. Exploitation grows too, but until spring 2026 it grew more slowly than disclosures: comparing the same ten months, the share of new CVEs that reach KEV fell from 0.54 % in 2023 to 0.23 % in 2026. Since May 2026, Google’s threat intelligence data shows exploitation growing in step with disclosures, and the time between disclosure and exploitation keeps shrinking. For defenders, a raw CVE count measures disclosure activity rather than risk, but the risky subset is now growing and arriving faster.

1 Introduction#

A CVE ID is a name for a publicly disclosed vulnerability. It is not a measure of danger, but CVE counts are routinely quoted as one, both in vendor reports and in budget discussions. The volume also matters operationally: every record has to be triaged by someone, and the US National Vulnerability Database (NVD) stopped keeping up in early 2024 [1].

While building cve.zaage.it, a tracker that combines the CVE List, KEV, EPSS, the NVD and the GitHub Advisory Database, I wanted to answer three questions with its data:

  1. Is the growth exponential?
  2. What causes it?
  3. Does exploitation grow at the same rate?

2 Data and method#

  • CVE records: the CVE Program’s cvelistV5 repository [2], snapshot of 29 October 2026, with 397,528 published and 18,843 rejected IDs. A record counts towards the year and month of its datePublished (UTC). Rejected records are excluded.
  • Assigner: the CNA that published the record (assignerShortName), as given in the record.
  • Exploitation evidence:
    • CISA KEV [3], 1,762 entries;
    • FIRST EPSS [4], scores of 28 October 2026.
    • I treat a CVE as likely exploited if it is in KEV or has an EPSS score of at least 0.1, that is, an estimated chance of at least 10 % of exploitation activity within 30 days.
  • Trend:
    • an ordinary least-squares fit of log(count) on year for 2017–2025 (exponential model);
    • a linear fit on the same years, for comparison;
    • 2026 is held out.

All numbers are reproducible from the tracker’s API and its daily database build.

3 Results#

3.1 The shape of the growth#

Line chart, logarithmic scale, 2010 to 2026. Published CVEs rise from 4,639 in 2010 to 14,642 in 2017, 48,151 in 2025 and 88,650 by 29 October 2026, far above the dashed 2017 to 2025 trend of 16 percent a year that predicts 49 thousand for 2026. The likely exploited subset stays between about 800 and 1,300 a year.
Figure 1. Published CVEs per year and the likely-exploited subset (in KEV or EPSS ≥ 0.1), logarithmic scale. The dashed line is the 2017–2025 exponential fit. 2026 runs to 29 October.
YearPublishedChangeCNAs publishingTop-5 CNA shareIn KEV
20166,4573760.6 %0.84 %
201714,642+127 %680.61 %
201917,3089960.8 %0.75 %
202224,99920147.3 %0.52 %
202328,847+15 %2620.58 %
202439,924+38 %30753.1 %0.40 %
202548,151+21 %36556.6 %0.41 %
2026 (to 29 Oct)88,650+84 %40045.8 %0.23 %

Table 1. Published CVEs per year (selected years). KEV share is the fraction of that year’s CVEs that are in KEV today; recent years have had less time to be added.

For 2017–2025, the exponential model fits better than the linear one: its residual sum of squares is about half as large, and the implied growth rate is 15.7 % a year. This is a description, not a model of the process. The curve is not smooth; it moves in steps, and published forecasts such as FIRST’s are statistical time-series models rather than explanations of the growth [9]. I found no peer-reviewed work that separates “more vulnerabilities exist” from “more vulnerabilities are reported”. The steps:

  • 2017 (+127 %): the CVE Program federated assignment to many more CNAs [5].
  • 2024 (+38 %): the Linux kernel became a CNA in February 2024 and has issued thousands of CVEs a year since [6].
  • 2026: the exponential trend predicts about 49,000 records for the whole year. The actual figure passed that in July.

Ten months in, 2026 is already 84 % above all of 2025. Counted by ID year instead of publication date, it is 80,999 records, still 68 % above 2025. The same jump is visible in independent counts from the NVD [7, 8]. FIRST’s forecast, published in February, put the median for 2026 at 59,427 [9]. In June they raised it to about 66,000 [10]; that figure was passed in September.

3.2 Who publishes#

Stacked area chart of published CVEs per month from January 2024 to October 2026. The total stays near four thousand a month until early 2026, then climbs to 15,849 in October 2026. GitHub, VulnCheck, Microsoft and the Linux kernel CNA make up about six thousand of that; all other CNAs together make up almost ten thousand.
Figure 2. Published CVEs per month since January 2024, by CNA. October 2026 runs to the 29th.
CNAJan–Oct 2025Jan–Oct 2026Factor
GitHub2,25612,2405.4
VulnCheck5609,07916.2
Linux kernel4,5168,2011.8
VulDB5,1516,1061.2
Microsoft1,1084,4204.0
Oracle3144,15413.2
Google Chrome1103,64633.1
Apache1791,2477.0
Red Hat2571,2514.9
Mozilla1768594.9

Table 2. CNAs with the largest increases, January–October 2026 against the same months of 2025.

No single CNA dominates the 2026 increase. The four CNAs named in Figure 2 together averaged 905 records a month in 2025 and about 6,100 in October 2026. All other CNAs grew from 3,108 to 9,729 a month. At the same time the market became less concentrated: the Herfindahl index across CNAs fell from 2,121 in 2019 to 579 in 2026.

3.3 What gets exploited#

Recent years have had less time to be added to KEV, so I compare like with like: CVEs published between 1 January and 29 October of each year, counted as exploited if they reached KEV within 300 days of publication.

YearPublished (1 Jan–29 Oct)In KEV within 300 daysShareChange in KEV count
202323,4411260.54 %
202432,2701170.36 %−7 %
202539,3331540.39 %+32 %
202688,6502010.23 %+31 %

Table 3. Like-for-like comparison of disclosures and KEV listings. KEV started in November 2021, so earlier years are not comparable.

Exploitation grows, but more slowly than disclosures: in 2026 the number of exploited new CVEs grew by about a third while the number of CVEs more than doubled. Two independent measurements agree on the direction for the first half of 2026:

  • VulnCheck counted 495 known exploited vulnerabilities in H1 2026 in its own catalogue, which is broader than CISA’s: 10 % more than in the previous six months, while CVE volume grew 45 %. VulnCheck calls it “too early” to tell whether exploitation will scale with issuance [22].
  • Google’s Threat Intelligence Group (GTIG) recorded 141 exploited vulnerabilities from January to August 2026, more than the 127 in all of 2025: 18 a month instead of 10.5 [8].

GTIG also finds that the gap is closing. In GTIG’s words: “Since May 2026, a shift has emerged, with the expansion of CVE exploitation (+127% indexed growth) closely mirroring disclosure growth (+128% indexed growth)” [8]. Exploitation of vulnerabilities GTIG rates as high-risk more than doubled, from 28 in 2025 to 75 in the first eight months of 2026, and most of the growth comes from n-days, vulnerabilities that are weaponised quickly after disclosure [8].

Exploitation also gets faster:

  • For CVEs published in 2026 that reached KEV, the median gap between publication and listing is 4 days, against 8 days for 2025 in the tracker’s data.
  • VulnCheck measures the median from publication to KEV falling from 120 days in 2025 to 80 days in H1 2026 [22]; the gap to my figure is probably one of populations: mine counts only 2026 CVEs that are already listed, which by construction were listed quickly.
  • Mandiant’s average time to exploit fell from 63 days in 2018 to −1 day in 2024, and an estimated −7 days in 2025: on average, exploitation now starts before a patch is available [11]. This is a mean over exploited vulnerabilities only, pulled negative by zero-days, not a figure for all CVEs.

4 Discussion#

4.1 Why the count grows#

The data identifies who publishes more; it does not by itself show why. Separating what is established from what is plausible:

Established

  • CNA structure explains the steps before 2026. The number of CNAs publishing per year grew from 37 (2016) to 400 (2026). The 2017 federation and individual CNA policies produce visible jumps.
    • WordPress-ecosystem CNAs (Patchstack, Wordfence, WPScan) and VulDB made thousands of plugin vulnerabilities per year visible from 2022 on.
    • The Linux kernel CNA assigns CVEs to nearly every fix that may be security-relevant [6]. GTIG notes about 5,000 kernel CVEs from January to August 2026 with no observed exploitation of a kernel zero-day [8].
  • Backfilling is a measurable but minor part of 2026. 7,651 of the 88,650 records published in 2026 (8.6 %) carry an ID from an earlier year, 1,996 of them from VulnCheck.
  • Large vendors attribute part of their 2026 volume to AI tooling. Google says AI helped Chrome fix 1,072 security bugs in Chrome 149 and 150, more than in the previous 23 milestones combined; it names Big Sleep and a Gemini-based agent harness [23]. Microsoft’s agentic scanner MDASH found 16 of May’s fixes, and Patch Tuesday grew from a then-record of more than 200 CVEs in June to 974 in September [13, 24].

Plausible, not proven

  • More software and more scrutiny explain the growth between the steps. Fuzzing at scale is one measurable source: OSS-Fuzz alone reports over 13,000 vulnerabilities fixed [14]. Its share of CVEs is not known.
  • AI-assisted discovery is the main driver of the 2026 jump. The timing fits and the vendors with the largest increases (Chrome, Oracle, Microsoft, Apache, Mozilla) run large, mature code bases, where automated review should find the most. VulnCheck measured +563 % for Chrome and +476 % for GitHub by mid-May, but cautions that “not all increases can be directly attributed to AI” [12].
    • The evidence is mostly self-attribution and timing. CVE records carry no field for how a vulnerability was found [8], and independent tracing is thin: by mid-April 2026, VulnCheck found only one CVE explicitly attributed to Anthropic’s Project Glasswing [25].
  • Part of the vendor surge may be a change in assignment practice rather than in discovery. Internally found bugs were often fixed without CVEs; if AI-found internal bugs now receive CVEs, the count rises without more vulnerabilities being found. I found no vendor statement confirming or excluding this.
  • Some of the 2026 volume is low quality. Rejections rose from 1,833 in 2025 to 2,497 so far in 2026, and maintainers report AI-generated reports of little value; curl ended its bug bounty in January 2026 for that reason [26]. How much of 2026 will be rejected later is not yet known.

Unknown

  • Whether the underlying number of vulnerabilities grows, or only the rate at which an existing stock is found. Many AI-found bugs are old (Glasswing’s examples include a 27-year-old OpenBSD flaw [25]), which points to faster discovery of existing flaws.
  • Whether commercial incentives among CNAs (vendors of vulnerability data competing on coverage) contribute.
  • I found no evidence for the EU Cyber Resilience Act as a driver: its reporting duty, in force since 11 September 2026, covers actively exploited vulnerabilities reported to CSIRTs and ENISA, not CVE assignment [15].

4.2 What it means#

  • CVE counts measure disclosure, not risk. In Jerry Gamblin’s words, the volume reflects “who is publishing rather than what is breaking” [7]. A vendor whose CVE count doubles may simply have started looking.
  • But the risky subset is growing, and arriving faster. Exploited CVEs per year are up by about a third in 2026, and by GTIG’s measure growing as fast as disclosures since May [8]. With a negative average time to exploit [11], waiting for KEV is no longer early enough for exposed systems.
  • Triage capacity does not scale with volume. In Cyentia’s 2019 data, organisations fixed about 10 % of their open vulnerabilities per month, regardless of size [16]; I found no newer measurement. With four times the input, the backlog grows unless prioritisation becomes stricter.
  • Severity alone is a poor filter. 54 % of 2026’s scored records are rated High or Critical. CVSS describes impact, not likelihood, and has well-known weaknesses as a prioritisation tool [17].
    • Exploitation evidence (KEV) and prediction (EPSS, which in its third version improved on its predecessor by 82 % [18]) target the few hundred CVEs a year that are actually used.
  • The shared infrastructure is under strain.
    • The NVD’s backlog grew to over 27,000 unanalysed CVEs by the end of 2025 [1].
    • Since April 2026 NIST has fully enriched only KEV entries and software used by the US government or listed as critical software. Older backlog entries are labelled “Not Scheduled” [19].
    • In September 2026 CISA announced a “Quality Era” for the CVE Program [20].
    • Tools that rely on NVD scores alone now see a growing share of unscored records.

5 Limitations#

  • Counting rules: independent tallies differ by up to about 2,000 records a year. Gamblin, for example, counts 48,185 for 2025 and this post 48,151 [7]. The differences come from the date used, the handling of rejected records, and counting by publication year or ID year.
  • Dates: the counts use the CVE List’s publication date. The NVD, and therefore most published statistics, uses its own date, which is why monthly figures differ between sources. For example, Google counts 10,740 for August 2026 and the tracker 12,259 [8].
  • EPSS is not comparable across years: a CVE’s score is low when it is new and rises as evidence accumulates. The likely-exploited subset for 2025 and 2026 is therefore underestimated, while KEV shares are lower bounds.
  • KEV covers only part of exploitation, and lags: it lists vulnerabilities with reliable public evidence of exploitation, mostly relevant to US federal systems, and CISA adds only about 190–280 entries a year. Listing often follows exploitation by weeks or months, so recent months are undercounted. GTIG’s measurements may be closer to current exploitation than KEV.
  • Attribution is partly secondary: the attribution of the 2026 surge to AI rests on vendor statements and press reports, not on fields in the records. Several vendor figures are taken from press coverage.
  • The trend fit is descriptive: with visible regime changes, a single exponential fit summarises growth but does not predict it.

6 Conclusion#

From 2017 to 2025, CVE counts grew at about 16 % a year, in steps that follow changes in who may assign CVEs. In 2026 the series left that trend: monthly volume roughly quadrupled, carried by major vendor CNAs, GitHub, VulnCheck and the kernel CNA, at the same time as AI-assisted discovery arrived. AI is the most likely driver, but the records cannot prove it, and part of the increase may be a change in what gets a CVE rather than in what gets found. Does exploitation follow? Partly. The number of exploited CVEs grows, by about a third in 2026, but until spring it grew far more slowly than disclosures, so the exploited share halved. Since May 2026 the two have been growing at the same rate, and exploitation starts sooner after disclosure every year. As the signal-to-noise ratio of a raw CVE feed falls [21], prioritisation must start from exploitation evidence, and then use EPSS and asset context, rather than from the count or from CVSS alone. As exploitation speeds up, that prioritisation also has to happen within days.

The live numbers behind this post are on cve.zaage.it.

References#

  1. U.S. Department of Commerce, Office of Inspector General. Evaluation of NIST’s Management of the National Vulnerability Database. Report OIG-26-020-I, 26 May 2026. oig.doc.gov
  2. CVE Program. cvelistV5. github.com/CVEProject/cvelistV5
  3. CISA. Known Exploited Vulnerabilities Catalog. cisa.gov
  4. FIRST. Exploit Prediction Scoring System (EPSS). first.org/epss
  5. CVE Program. CVE Numbering Authorities. cve.org
  6. P. Przymus et al. Linux Kernel Recency Matters, CVE Severity Doesn’t, and History Fades. MSR ‘26, arXiv:2601.22196, January 2026. arxiv.org
  7. J. Gamblin. 2025 CVE Data Review, 1 January 2026, jerrygamblin.com; and monthlyCVEStats, September 2026, github.com
  8. R. Grunewald, S. Mazumdar, K. Vanderlee (Google Threat Intelligence Group). Vulnerability Discovery and Exploitation Trends in the AI Era. 30 September 2026. cloud.google.com
  9. FIRST. Vulnerability forecast 2026: The Year Ahead, February 2026. first.org
  10. M. Zorz. Report on FIRST’s updated 2026 CVE forecast. Help Net Security, 15 June 2026. helpnetsecurity.com
  11. Mandiant. M-Trends 2026, as reported by Help Net Security, 24 March 2026. helpnetsecurity.com
  12. P. Garrity. The First CVE Wave: Signs That AI-Assisted Vulnerability Discovery Is Reshaping Disclosure Volumes. VulnCheck, 14 May 2026. vulncheck.com
  13. J. Coker. Microsoft Shatters Patch Tuesday Record With 974 CVE Fixes in September 2026. Infosecurity Magazine, 9 September 2026. infosecurity-magazine.com
  14. Google. OSS-Fuzz. github.com/google/oss-fuzz
  15. McCann FitzGerald. Cyber Resilience Act: reporting obligations apply from 11 September 2026. mccannfitzgerald.com
  16. Cyentia Institute, Kenna Security. Prioritization to Prediction, Volume 3. cyentia.com
  17. J. Spring, E. Hatleback, A. Householder, A. Manion, D. Shick. Time to Change the CVSS? IEEE Security & Privacy 19(2), 2021. doi:10.1109/MSEC.2020.3044475
  18. J. Jacobs, S. Romanosky, O. Suciu, B. Edwards, A. Sarabi. Enhancing Vulnerability Prioritization: Data-Driven Exploit Predictions with Community-Driven Insights. 2023. arXiv:2302.14172
  19. NIST. Announcement on prioritising NVD enrichment, 15 April 2026. nist.gov
  20. CISA. CVE Program: Establishing a Quality Era Framework, September 2026. govdelivery.com
  21. A. Wahid. Rising CVEs in the AI epoch. APNIC Blog, 1 July 2026. blog.apnic.net
  22. P. Garrity. VulnCheck State of Exploitation 1H-2026. VulnCheck, 28 July 2026. vulncheck.com
  23. L. Abrams. Google says AI helped Chrome fix 1,072 security bugs in two releases. BleepingComputer, 30 July 2026. bleepingcomputer.com
  24. Microsoft patches record 200-plus vulnerabilities as AI accelerates bug discovery. SiliconANGLE, 10 June 2026. siliconangle.com
  25. S. Sharma. Behind the Mythos hype, Glasswing has just one confirmed CVE. CSO Online, 16 April 2026. csoonline.com
  26. S. Sharwood. Curl shutters bug bounty program to remove incentive for submitting AI slop. The Register, 21 January 2026. theregister.com

cd ~/writing