The number of Common Vulnerabilities and Exposures published in 2026 passed 100,000 on Friday, a record, and more than double the whole of last year — with seven weeks still left in the year.

Vulnerability intelligence firm VulnCheck counted 100,222 CVEs at around 16:00 UTC on Friday, according to The Register. The comparison is stark: 2025 closed with roughly 48,185 published CVEs, and the year before that with about 40,000. In 2026 the industry passed the previous annual record before the end of August.

The surge is real but it is not evenly explained. Some of it is ordinary: more software, more vendors, more mandatory disclosure. Some is structural: Microsoft alone has shipped record Patch Tuesday volumes this year, 1,414 flaws in October after a thousand-plus in September, and each of those becomes one or more CVE records. And some of it is new machinery — AI models pointed at source code.

The most consequential test case is Anthropic's Project Glasswing, launched on April 7 with a claim of having found "thousands of additional high- and critical-severity vulnerabilities" in open-source software. VulnCheck has been auditing that claim from the outside since April, tracking every CVE, GitHub Security Advisory and kernel commit it can attribute to Anthropic.

Seven months in, its tally is 442 CVEs attributable to Anthropic — about 0.5% of everything published this year, or roughly 63 a month. Of those, only two have been exploited in the wild: CVE-2026-61500 and CVE-2026-26980, or 0.45% of the set. Anthropic's own public ledger, updated in bulk on November 2, reports 37,966 findings, of which 7,349 (19.4%) have been made public and just 217 have been marked fixed. Since September the project added nearly 12,000 findings and 15 fixes. Duplicate reports reached 1,654 (4.4%) — nearly eight times the number fixed — and 419 findings were withdrawn, even as the claimed true-positive rate rose from 90.8% to 94.2%.

The clearest evidence comes from a maintainer rather than a vendor. Daniel Stenberg of curl received a batch of 45 reports from Anthropic; of the first 34 published, 19 were marked not-a-vulnerability, six confirmed as low severity, and nine remain unassessed. That is a 55.9% false-positive rate so far, with nothing confirmed above low severity, and Stenberg is publishing them at a deliberate pace "to soften the impact for the team". Anthropic's response has been to hand triage back: its opt-in OSS Scanner, launched October 8, sends model-generated reports with no human review and warns they may be invalid.

None of this means AI is bad at finding bugs. VulnCheck's own first-half review counted 1,061 vulnerabilities attributed to AI-assisted discovery, and 14 of them — 1.3% — had been exploited. That is a worse exploitation ratio than the CVE population as a whole suggests is possible, not a better one. What has changed is the ratio of discovery to verification. Finding a candidate bug is now cheap and parallel; deciding whether it is real, severe and worth a maintainer's week is still human, serial and slow.

The practical consequence is the number that matters least: 100,000. Vulnerability management does not run on totals, it runs on prioritisation. CISA's Known Exploited Vulnerabilities catalogue added 31 entries between October 20 and November 13 — a workload an administrator can actually finish, unlike a five-figure annual count. The 100,000th CVE will not be the one that hurts anyone; the handful of entries added to that list almost certainly will.

The other consequence is quieter and possibly more damaging. If a meaningful share of the new volume is machine-generated and wrong, the cost lands on volunteer maintainers who must read it, and on the credibility of a disclosure system already strained by the National Vulnerability Database's backlog of unanalysed records. A registry that grows faster than its triage capacity stops being an index of risk and becomes a queue.