The CVE program logged more than 40,000 entries in 2024. In April 2026, NIST decided its National Vulnerability Database would reserve detailed analysis for selected risks, even as AI systems began producing vulnerability findings and proposed fixes at scale.

Key takeaways

  • In 2024, 413 organizations reported more than 40,000 CVEs, bringing the cumulative CVE record above 270,000.
  • In March 2024, NIST analyzed 199 of the 3,370 CVEs it received that month.
  • NIST’s April 2026 NVD policy reserves detailed analysis for actively exploited vulnerabilities, critical software, and systems used by the federal government.
  • OpenAI reported validating and patching 948 Linux local-privilege-escalation vulnerabilities for approximately $1.2 million in API costs.
  • In February 2024, CISA ordered federal agencies to disconnect vulnerable Ivanti VPN appliances before a deadline because the flaws were under active exploitation.

That changes the job around CVE. Its public identifier still lets researchers, vendors, agencies and defenders refer to the same flaw. But as AI makes candidate findings—and sometimes proposed fixes—cheaper to produce, the scarce service is deciding which named flaws merit further analysis and who must act on them. NIST’s April decision makes that division explicit: CVE remains the common name; enrichment becomes selective.

CVE scaled naming without promising equal analysis

MITRE’s CVE program began in 1999 to solve a practical coordination problem: one vendor’s advisory and another defender’s alert needed a common reference. The program did not need a single desk to write every entry. In 2024, 413 organizations reported more than 40,000 CVEs, bringing the cumulative record above 270,000. That distributed assignment model is a strength, not evidence that naming has failed.

The National Vulnerability Database performs a different task. Built on the CVE list, NVD adds analysis that helps security teams sort and automate their work. An assigned identifier establishes what people are discussing; it does not establish whether an attacker can exploit the flaw, whether a particular organization runs the affected software, or whether a fix is safe to deploy. The distinction matters more when the list grows faster than the capacity to examine it.

NIST put an admission rule on the scarce work

In March 2024, NIST analyzed 199 of the 3,370 CVEs it received, according to SC Media’s coverage of its backlog. That is a snapshot of one month’s intake, not a measure of CVE assignment or NIST’s permanent analysis rate. By July 2025, reporting put the NVD backlog above 25,000 vulnerabilities—roughly ten times its previous 2017 high.

In April 2026, NIST chose where its analysts’ attention would go: vulnerabilities under active exploitation, in critical software, or in systems used by the federal government, including entries in CISA’s Known Exploited Vulnerabilities catalog. Reporting linked the backlog NIST was addressing to a 2024 funding lapse. The two-speed vulnerability-data system is therefore not just a response to AI. Public analysis capacity was already constrained before this year’s reported AI-assisted discovery volumes.

NIST’s rule also has a boundary. Federal use and critical software are public-interest priorities; they cannot tell every hospital, manufacturer or software maintainer which flaws affect its own deployed systems. Each operator still has to connect the shared identifier to its inventory and its ability to remediate.

A generated fix does not prove deployment

AI can relieve some of the work that follows discovery. OpenAI reported that a pipeline using an upcoming cyber model validated and patched 948 Linux local-privilege-escalation vulnerabilities for approximately $1.2 million in API costs. Microsoft has introduced Project Perception as an agentic system designed to patch vulnerabilities. Neither claim should be reduced to “AI only creates more alerts.” OpenAI’s API cost, however, is not a total remediation cost, and the cited record does not establish where those fixes were deployed or how they performed there.

The input side is less orderly. Anthropic reported 5,500 verified vulnerabilities found from April through October; Glasswing partners reported more than 129,000 findings from April through July, including more than 33,000 rated critical or high severity. Anthropic and Glasswing partners reported separate totals without showing whether their findings overlap or how many became CVEs, received NVD analysis, or reached deployed fixes. Meanwhile, Google paused product-flaw submissions to its OSS Vulnerability Reward Program amid invalid AI-driven reports. A report generator can lower the cost of submission for everyone, including people who have not done the verification.

Maintainer capacity is not uniformly slow, either. Project Zero reported an average of 25 days for Linux developers to patch security holes, versus 44 for Google, 69 for Apple and 83 for Microsoft. A credible finding can move quickly through some ecosystems. NVD cannot infer that capacity from the identifier, and an AI-produced patch cannot assign responsibility for accepting it.

CISA shows how a name becomes an instruction

CISA’s Known Exploited Vulnerabilities catalog adds a consequential fact to a shared identifier: attackers are known to be using the flaw. CISA can then require federal agencies to remediate within specified timeframes. In February 2024, it went further, ordering agencies to disconnect vulnerable Ivanti VPN appliances before a deadline because the flaws were actively exploited. That order depended on more than a severity label. CISA identified an affected class of deployed equipment, an active threat and an agency able to take action.

Frequently asked questions

Which Linux vulnerabilities and deployments were covered by OpenAI’s 948-patch result?

The piece identifies them only as Linux local-privilege-escalation vulnerabilities. It does not name the individual flaws, affected distributions, organizations that deployed fixes, or post-deployment results.

How will NIST rank vulnerabilities that all meet its new analysis criteria?

The piece states the qualifying categories—active exploitation, critical software, and federal systems—but does not provide a ranking formula, service-level target, or expected number of detailed analyses.

Did Google resume product-flaw submissions to its OSS Vulnerability Reward Program after pausing them?

The piece says Google paused those submissions amid invalid AI-driven reports. It gives no resumption date, revised submission rules, or outcome for the paused reports.

What exact remediation deadlines apply when CISA adds a flaw to the Known Exploited Vulnerabilities catalog?

The piece says CISA can require federal agencies to remediate within specified timeframes and cites an Ivanti disconnection order in February 2024. It does not provide the catalog’s specific deadlines or a deadline for that order.

From CVE scale to selective NVD analysis

  • 1999 — MITRE’s CVE program began to provide a common reference for vulnerabilities.
  • March 2024 — NIST analyzed 199 of the 3,370 CVEs it received that month.
  • 2024 — 413 organizations reported more than 40,000 CVEs, pushing the cumulative record above 270,000.
  • July 2025 — Reporting put the NVD backlog above 25,000 vulnerabilities, roughly ten times its previous 2017 high.
  • April 2026 — NIST limited detailed NVD analysis to actively exploited vulnerabilities, critical software, and systems used by the federal government.

The 40,000-plus CVEs logged in 2024 still give defenders common names. NIST decides which entries receive detailed analysis; CISA can order action on exploited flaws, while operators must match those names to their own systems before they know what to fix.