Field brief · vulnerability data · November 2026
NVD published 49,114 high or critical CVEs through 6 November 2026. That is nearly every high-severity CVE from 2016 through 2021, compressed into one year to date. AI-assisted discovery is part of the reason; automated CNA assignment is more of it; exploitation is not scaling with either.
From 2016 through 2021, high and critical disclosures moved in a band between roughly 3,800 and 10,300 a year. The line bent in 2022, crossed 19,500 in 2024, hit 23,321 in 2025, and then went vertical in 2026. The chart is stacked: blue is CVSS high (7.0–8.9), red is critical (9.0–10.0).
Source: NVD CVE JSON 2.0 annual feeds, pulled 6 November 2026, grouped by NVD published date. 2026 is year-to-date through 6 November. Severity is the highest-priority CVSS base severity on each record (NVD primary preferred; CNA metrics used where NVD has not scored).
2016 high + critical
3,765
58.0% of all CVEs
2025 high + critical
23,321
46.8% of all CVEs
2026 YTD
49,114
51.8% of all CVEs
2026 annualized
~57,800
if the current rate holds
| Year | All CVEs | High | Critical | High + critical | Share |
|---|---|---|---|---|---|
| 2016 | 6,487 | 2,882 | 883 | 3,765 | 58.0% |
| 2017 | 16,669 | 6,569 | 2,076 | 8,645 | 51.9% |
| 2018 | 18,044 | 7,507 | 2,583 | 10,090 | 55.9% |
| 2019 | 18,363 | 7,023 | 2,510 | 9,533 | 51.9% |
| 2020 | 18,641 | 7,297 | 2,576 | 9,873 | 53.0% |
| 2021 | 21,848 | 7,814 | 2,531 | 10,345 | 47.3% |
| 2022 | 26,378 | 9,880 | 4,194 | 14,074 | 53.4% |
| 2023 | 30,716 | 10,948 | 4,608 | 15,556 | 50.6% |
| 2024 | 40,698 | 14,516 | 5,011 | 19,527 | 48.0% |
| 2025 | 49,802 | 16,929 | 6,392 | 23,321 | 46.8% |
| 2026 YTD | 94,750 | 38,173 | 10,941 | 49,114 | 51.8% |
2026 covers 1 January–6 November. Shares are high + critical divided by all CVEs with a published date in the window. The 2016 baseline is affected by CVSS v3 adoption and NVD coverage; the post-2021 trend is not.
The shape matters more than any single year. This is not a smooth 5% annual drift. From 2021 to 2025 the high-severity count rose 2.25×. The 2026 year-to-date count alone is nearly twice the entire 2021–2025 increase. Something in the discovery and reporting pipeline changed, not just the amount of software in the world.
The monthly series is where the acceleration is unmistakable. November 2025 was the trough at 1,465 high or critical CVEs. October 2026 was 8,526. The 2025 monthly average was 1,943; October 2026 is 4.4× that. The last twelve months total 52,124, against 22,940 in the prior twelve.
Same source and severity rule as the annual chart. Months are NVD published months. November 2026 is excluded because the month is incomplete.
| Month | High | Critical | High + critical | All CVEs |
|---|---|---|---|---|
| Nov 2025 | 1,070 | 395 | 1,465 | 3,115 |
| Dec 2025 | 1,952 | 592 | 2,544 | 5,642 |
| Jan 2026 | 1,696 | 574 | 2,270 | 5,141 |
| Feb 2026 | 1,748 | 587 | 2,335 | 4,808 |
| Mar 2026 | 2,608 | 862 | 3,470 | 6,300 |
| Apr 2026 | 2,239 | 653 | 2,892 | 5,871 |
| May 2026 | 2,924 | 731 | 3,655 | 6,984 |
| Jun 2026 | 3,432 | 984 | 4,416 | 8,000 |
| Jul 2026 | 4,234 | 1,384 | 5,618 | 9,918 |
| Aug 2026 | 5,300 | 1,822 | 7,122 | 12,719 |
| Sep 2026 | 6,267 | 1,544 | 7,811 | 15,073 |
| Oct 2026 | 6,911 | 1,615 | 8,526 | 17,470 |
The step changes are not random. March 2026 is the first month above 3,000. June crosses 4,000. August crosses 7,000. September and October add another 700 each. That is the signature of capacity being added to the discovery pipeline, not a one-off vendor patch day — though vendor patch days are also in the numbers.
Four things moved at once. They are not equally responsible, and separating them is the whole argument.
This is no longer a demo. Mozilla's Q1 2026 CVE volume rose 164%, which FIRST attributes directly to AI-assisted tooling running against the Firefox engine. Google's threat-intelligence group tracked a distinct population of AI-discovered CVEs and found they skew toward real impact: 50% result in remote code execution, versus 26% across the broader ecosystem; only 39% are low-risk, versus 69% for conventional discovery. Anthropic's Project Glasswing reported more than 6,000 critical or high-severity findings across open-source projects.
This is the part most “CVE explosion” posts miss. GitHub Security Advisories are up 449% year over year. VulnCheck's CNA-of-last-resort activity is up 3,119%, absorbing a backlog of unassigned issues. In my pull, 8,266 CVEs published in 2026 — 8.7% of the year's records — have a description that mentions “Linux kernel.” Google counted roughly 5,000 Linux-kernel CVEs from January to August 2026 and zero observed in-the-wild zero-days among them. A CVE is now often one advisory per patch, not one advisory per incident.
FIRST's forecast team notes that the number of distinct software products with tracked vulnerabilities has grown by two orders of magnitude. More packages, more cloud services, more firmware, more AI middleware. Google tracked 2,076 AI-stack CVEs from January 2025 through August 2026, more than 1,500 of them in 2026 alone: 782 in agent orchestration frameworks, 212 in inference and serving infrastructure, and 24% of the serving bugs traceable to unauthenticated APIs or SSRF.
Bug bounties, vendor patch days, and diffable advisories were already industrial. AI collapsed the cost of the next step: reading the patch, finding the trigger, and writing a working exploit. Google measured exploitation rising from 10.5/month in 2025 to 18/month in January–August 2026, while zero-day exploitation moved only from 8/month to 11/month. The growth is mostly n-days, not new zero-days. That is the signature of faster weaponization, not faster discovery alone.
The interesting question is not “can a model find a bug?” It can. The question is which techniques turn a model into a repeatable discovery pipeline, and which of those techniques are now available outside a frontier lab.
Take a known CVE, extract the bug class, and ask an agent to find the same pattern in adjacent code or a different product. Hacktron found the BeyondTrust pre-authentication RCE this way: its agents ran variant analysis on the bug class from an earlier Ivanti EPMM RCE and scanned enterprise software for the same shape. Google's agentic vulnerability-discovery harness does the same thing against client codebases. This is the highest-leverage use of an LLM in offense right now because it converts one public disclosure into a search query over every similar codebase.
The useful systems are agents with tools: repository search, build, test, debugger, and a loop. They read across files, follow data flow, write a proof harness, run it, and revise. Google's own framing is that these systems excel at memory corruption and logic bypasses in C/C++ libraries and runtimes — the exact classes that produce RCE. That matches the observed skew: half of AI-discovered CVEs are RCE.
A model that can write the fuzzer, compile it, read the crash, and minimize the input is doing the expensive part of vulnerability research. It does not need to be right every time; it needs to be cheap enough to run thousands of times. That is why the volume effect shows up even when per-attempt accuracy is modest.
Once a patch ships, the diff is public. LLMs are good at reading the diff, inferring the pre-patch trigger, and generating a proof of concept. Google's read is that threat actors are using LLMs to weaponize n-days faster, which fits the pattern: exploitation grows with disclosure while zero-day counts stay flat. The bottleneck moved from “find a bug” to “weaponize a published fix,” and that bottleneck is exactly where models are strongest.
This is the part the user thesis gets right, with one caveat. Open-weight models are now close enough on vulnerability-discovery tasks to matter, and they have no per-token bill, no provider rate limit, and no central policy switch. Edera's mid-2026 comparison put the top open-weight model at 16 of 26 known vulnerabilities found, against 23 for the top-performing proprietary model (Grok); multiple passes narrow the gap further. On AgenticVulHunter's secure-code-review benchmark, a locally deployed Qwen3-Coder-30B scored 18.4% and DeepSeek V3.1 Terminus scored 23.8%, against 17.5% for AgenticSCR using Claude Sonnet 4. The absolute accuracy is low; the point is that the marginal review pass is now free enough to run on every commit, every package, and every appliance firmware image.
What open weights do not do is remove the need for a harness, a target, and human triage. The 2026 numbers are consistent with a world where many more codebases get a first-pass review, not a world where a laptop autonomously produces zero-days. The capability gap that mattered was never “can it find one?” It was “can you afford to run it everywhere?” That gap narrowed sharply.
I ranked by CVSS, confirmed exploitation, and blast radius, not by raw score alone. All three are remotely reachable, require no credentials, and were exploited within days. Microsoft's 10.0 Entra ID bug (CVE-2026-69836) is not on the list: Microsoft retracted the exploitation claim and says no customer action is required.
| # | CVE | Product | CVSS | Class | Exploitation |
|---|---|---|---|---|---|
| 1 | CVE-2026-85706 | GitLab CE/EE | 10.0 | Path traversal + missing auth → unauth arbitrary file read | Probes within hours; CISA KEV 11 Sep 2026 |
| 2 | CVE-2026-58231 | SAP Commerce Cloud (Data Hub Adapter) | 10.0 | Default auth client abuse + input validation → unauth RCE | Exploited three days after disclosure |
| 3 | CVE-2026-1731 | BeyondTrust RS / PRA | 9.8 / 9.9 | Pre-auth OS command injection | Exploited within four days; AI-discovered |
CVSS 3.1 base scores from NVD; BeyondTrust is 9.8 on v3.1 and 9.9 on v4.0. “Exploited” means confirmed in-the-wild activity, not a proof of concept.
CVSS 10.0 · unauth arbitrary file read · CISA KEV
The repository commits API built a file path from user input and did not confine it to the repository. The same endpoint also failed to enforce authentication. A crafted request could traverse out of the repo and read files the GitLab process could see — gitlab.rb, logs, CI variables, and SSH material. GitLab's vector is AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:N: maximum confidentiality and integrity impact, scope changed, no availability impact.
The practical exploit is credential theft, not a direct shell. watchTowr observed probes from 06:00 UTC on 11 September 2026 and later saw attackers dumping config files and system SSH configuration. The only precondition is at least one public project on the instance, which many exposed instances have. CISA added it to KEV the same day; federal remediation was due 14 September. If you run self-managed GitLab and it touches the internet, this is the one to check first.
CVSS 10.0 · unauth RCE · exploited in three days
SAP Commerce Cloud's Data Hub Adapter exposed a default authentication client that an unauthenticated attacker could abuse to reach functions lacking sufficient authorization and input validation. The result was arbitrary code execution and compromise of internal components. The scope is changed (S:C), so the blast radius is not limited to the adapter itself.
SAP shipped the fix in its 11 August 2026 Patch Day. Exploitation attempts followed within three days. This is the cleanest example of the 2026 pattern: a maximum-severity, no-credentials, network-reachable bug in a business-critical system, with a public patch diff and a very short window between “patch available” and “someone is trying.”
CVSS 9.8 / 9.9 · pre-auth command injection · AI-discovered
BeyondTrust Remote Support and older Privileged Remote Access passed attacker-controlled request data into OS command execution before authentication. A crafted request could run commands as the site user on an appliance that brokers privileged sessions — a foothold with a direct path to credentials, managed endpoints, and lateral movement. Roughly 11,000 instances were internet-exposed, about 8,500 of them on-premises and therefore dependent on the customer patching.
Hacktron's agents found it through variant analysis of an earlier Ivanti EPMM RCE. BeyondTrust confirmed the report the same day and patched SaaS within two days. Google's threat-intelligence group observed one cluster exploiting it within four days of disclosure, and five more within seven, followed by privilege escalation, data exfiltration, and secondary payloads. This is the case that best supports the AI thesis: a real, high-impact bug, found by an autonomous research pipeline, exploited on a compressed timeline.
The honest summary is “rain, not flood” for risk and “flood” for paperwork. FIRST's forecast team uses the same distinction: total CVE volume is up sharply, but vulnerabilities that are actively exploited or credibly exploitable have not risen at the same rate. Google's numbers make the gap concrete: 141 exploited vulnerabilities from January to August 2026, against 127 for all of 2025. Google puts the exploited share at 0.23% of 2026 disclosures — about 1 in 431.
2026 disclosures
94,750
through 6 Nov
Exploited Jan–Aug
141
vs 127 in all of 2025
Exploited share
0.23%
Google's Jan–Aug figure; about 1 in 431
That does not make the high-severity count meaningless. It means the count is an input to triage, not a risk score. The 2026 data supports four concrete moves:
published field. A CVE can be reserved, disclosed, and published on three different dates. NVD's enrichment backlog also means some records are published before they are fully analyzed; I use CNA metrics where NVD has not scored.