CVE Statistics 2026: 86,533 Records So Far and What Is Driving Them

The National Vulnerability Database (NVD) lists 86,533 CVE records published between 1 January and 25 October 2026, with rejected records left out. All of 2025 had 48,151. At about 290 records a day, 2026 is running at more than twice last year’s pace of about 132. We pulled the numbers on 26 October 2026.

In our assessment, three things push the count up and a fourth makes the extra flaws more urgent. The evidence is stronger for some than for others, and this post keeps what the sources state apart from our own assessment.

How Many CVEs Have Been Published in 2026 So Far?

Every year in the chart is counted the same way: NVD records by published date, rejected records excluded. The CVE Program’s own quarterly totals agree within about one percent for 2023, 2024, 2025 and the first half of 2026.

The growth did not arrive evenly. The CVE Program counts 15,163 records for January to March 2026, which is 26 percent above the same quarter of 2025. April to June had 20,709, up 77 percent. In our NVD pull, July to September reached 36,961, about 3.2 times the same months of 2025.

September 2026 alone had 14,933 records, against 4,321 in September 2025. October had 13,732 by the 25th, about 549 a day.

The count jumped in March, dipped in April and then rose every month from May to September. Anthropic announced Project Glasswing on 7 April (see below). The rise fits our assessment, but it began before that date, and NVD data does not say how a bug was found, so we do not call it proven.

Are Exploited Flaws Rising at the Same Speed?

No. CISA’s Known Exploited Vulnerabilities (KEV) catalog added 187 entries in 2023, 186 in 2024, 245 in 2025 and 272 in 2026 through 23 October. That is 11 percent above all of 2025, while the CVE count is 80 percent above it.

Per 1,000 new CVE records, KEV additions fell from 6.5 in 2023 to 3.1 in 2026. In our assessment, most of the extra records are not being exploited in the wild, at least not yet. The risk sits in a small slice, and that slice is still growing.

Is AI Finding the Extra Bugs?

In part, by the vendors’ own accounts. On 7 April 2026 Anthropic said Claude Mythos Preview “has already found thousands of high-severity vulnerabilities”. Its examples include a 27-year-old OpenBSD flaw and a 16-year-old FFmpeg flaw, which it says are now patched. Read the full statement on the Project Glasswing page.

The earlier evidence was DARPA’s AI Cyber Challenge. Its August 2025 results say the competing systems found 54 synthetic vulnerabilities and patched 43. They also found 18 real ones and produced 11 patches, across more than 54 million lines of code.

Vendors now do the same on their own products. In a post dated 21 October 2026, Fortinet says it “has integrated frontier AI models into our product security testing process”. It says this helped surface vulnerabilities more efficiently than conventional methods. Its hardening releases cover FortiOS, FortiManager, FortiAnalyzer, FortiMail and FortiAuthenticator, with public advisories due on Patch Tuesday, 10 November.

Our assessment: coordinated disclosure turns finds like these into CVE records over the following weeks and months, so AI-assisted research is a plausible driver of the 2026 curve. Neither source says what share of 2026 CVEs came this way. For a case where the cyber skills of an AI agent went past its test, see our coverage of the Mythos 5 test incident.

Do More Reporters and More AI-Written Code Add to the Count?

More reporters, clearly. The list of CVE Numbering Authorities (CNAs), the organizations allowed to assign CVE ids, held 263 entries at the end of December 2022. It held 346 in December 2023, 433 in December 2024, 484 in December 2025 and 559 on 20 October 2026. More CNAs mean more places that publish records.

The Linux kernel team shows the effect. Its documentation says it assigns CVE numbers to “any bugfix that they identify”, and gives that as the reason for the “seemingly large number” of kernel CVEs.

On the code side, the 2026 Stack Overflow Developer Survey (6 October 2026) had more than 30,000 respondents. It reports that 73 percent of those who use AI coding assistants or agents use them daily, and that 66 percent of AI users use coding agents. In our assessment, faster code output means more places for bugs to hide. The survey measures tool use, not the flaws that follow.

CISA published its Quality Era framework for the CVE Program on 22 September 2026. Its announcement said automated and AI-enabled technologies are introducing “new pressures across the software development lifecycle”.

Are Attackers Using the Same Tools?

Google’s Threat Intelligence Group (GTIG) reported on 11 May 2026 that, for the first time, it had identified a threat actor using a zero-day exploit. GTIG believes the exploit was developed with AI. The exploit was a two-factor authentication bypass in a popular open-source administration tool. GTIG also saw particular interest in AI-assisted vulnerability discovery from clusters linked to China and North Korea.

The UK National Cyber Security Centre judged in May 2025 that the time between disclosure and exploitation “has shrunk to days”, and that AI will almost certainly shorten it. CISA’s directive BOD 26-04, issued 10 June 2026, makes the same point about the time between patch release and exploitation.

Two 2026 flaws we covered show the pace defenders face: PHP2Shell (CVE-2026-109618 and CVE-2026-109619) and Node2Shell (a Node.js Brotli flaw). Our PHP2Shell timeline follows the first 60 hours after disclosure.

What Should Security Teams Do When 14,000 CVEs Arrive in a Month?

Do not try to read them all. BOD 26-04 gives a usable filter. It is written for US federal agencies, but CISA’s press release strongly encourages all partners to adopt similar actions. It asks four questions:

• Is the vulnerable asset publicly exposed?
• Is the CVE in the KEV catalog?
• Can an adversary automate every step of the exploit?
• Does exploitation give partial or total control?

CISA’s implementation guidance gives one example. A KEV-listed flaw on a publicly exposed asset, with total impact and automatable exploitation, gets a 3-day patching deadline.

In practice, we suggest this order:

1. Map what is reachable from the internet. The same flaw matters far more on an exposed host. Our External Attack Surface Management 101 guide explains how.
2. Check new CVEs against KEV first. With 272 additions in 2026 so far, that is fewer than one a day.
3. Add a probability signal. FIRST describes EPSS as a model that estimates the chance a published CVE will be exploited in the wild in the next 30 days. See how CVSS, EPSS and KEV signals compare.
4. Track days from fix release to fix deployed on exposed assets, and report that number.
5. Re-check your external surface whenever a widely exploited CVE appears.

For a longer treatment of the same idea, read our four filters that replace volume-based triage.

Brandefense External Attack Surface Management scans an organization’s internet-facing assets within minutes. To see what an attacker sees on your own domain, try a free check in Atlas, or request a demo. For the exploitation side, read our H1 2026 exploitation trends and the Vulnerability Management 101 guide.

How We Counted

NVD counts come from the NVD CVE API 2.0, filtered by published date with rejected records excluded. We pulled them on 26 October 2026 in windows of at most 90 days. The 2026 window ends on 25 October.

KEV counts come from CISA’s catalog JSON (version 2026.10.23, 1,756 entries), grouped by the year each entry was added. The KEV-per-1,000 figure divides one year’s additions by that year’s CVEs, so it is a rough ratio and not a rate for one cohort. CNA counts are the entries in the CVE Program’s published CNA list file, taken at the last update before each date.