NIST Analyzed 199 of 3,370 CVEs Before AI Scaled Up

Claroty and CISA had flagged critical vulnerabilities in three refrigerator-controller models. Attackers still used those flaws against supermarkets. The warning existed; the vulnerable machines remained in reach.

AI agents are adding candidate findings while NIST narrows the CVEs its National Vulnerability Database analyzes. Between a public warning and a completed repair, suppliers and operators still have to find the exposed machine.

NIST fell behind before AI agents multiplied findings

In March 2024, NIST analyzed only a fraction of the CVEs it received, according to contemporaneous coverage:

That shortfall predates the agents now producing more candidate work. OpenAI designed Codex Security to find vulnerabilities, validate them, and propose fixes. Anthropic said its Glasswing partners found more than 129,000 vulnerabilities from April through July 2026, including more than 33,000 rated critical or high severity. Those partner-reported findings are not a count of new CVEs, nor proof that each flaw was exploitable in a deployed system. They show how much candidate work a scanning network can produce without establishing what a particular customer must do.

In April 2026, NIST narrowed NVD analysis toward vulnerabilities in CISA’s Known Exploited Vulnerabilities catalog, critical software, federal systems, and issues under active exploitation. Its backlog followed a 2024 funding lapse. Other flaws remain in the CVE system, but NIST no longer commits to exhaustive public analysis.

The CVE program supplies an address, not an asset inventory

MITRE runs the CVE Program, whose identifiers let a supplier, a researcher, CISA, and a customer refer to the same disclosed issue. A 2025 account described 413 organizations reporting CVEs. Identifier assignment is a distributed undertaking, not one team at MITRE inspecting every affected installation. When MITRE said its funding had expired that year, CISA extended it and said critical CVE services would not lapse. The government protected the common naming layer.

That layer answers a necessary, limited question: which issue are these institutions discussing? A supplier must still establish affected products and versions. An enterprise team must determine whether those versions run on its assets, whether an exposed service can be reached, and who can approve a change. The CVE identifier allows each actor to join its evidence to the same issue; it cannot supply evidence that only the supplier or operator possesses.

CISA attached deadlines where NIST attached priority

NIST’s narrower approach has a defensible basis. Research on vulnerabilities disclosed between 2009 and 2018 found 4,183 exploited flaws among 76,000 discovered. That historical sample cannot identify which 2026 flaw will be exploited next, but it does show why equal analysis time for every record is a poor use of finite capacity.

CISA’s Known Exploited Vulnerabilities catalog makes one risk distinction operational: it identifies exploitation and sets remediation timeframes for federal agencies. NIST now gives such issues priority in its own analysis. CISA supplies a deadline to a defined set of operators; NIST concentrates public analytical work where that and other risk signals point.

A company that needs timely detail about a flaw outside NIST’s priority groups cannot assume a completed NVD analysis will arrive before it must decide. It may need the supplier’s advisory and its own product and asset records to establish exposure. NIST can direct limited capacity toward exploited flaws while leaving that broader information need unmet.

An automated patch still needs an accountable owner

Microsoft introduced Project Perception as an agentic system that patches vulnerabilities; OpenAI’s Codex Security proposes fixes as well as findings. Those products could remove work from the receiving end of the process, not merely add reports to it. A developer’s ability to reproduce and fix a bug already improves on a plausible claim that a maintainer must investigate from scratch—the problem explored in vulnerability-report triage.

A proposed code change and a completed repair remain different artifacts. A supplier can test a patch against its product. A customer must locate the affected deployment, schedule the change, check that it took, and retain a record of the result. Someone must accept responsibility for that chain, especially when a change might interrupt a working service. An agent can shorten the steps; the customer still needs proof that the patch reached the machine.

Attackers later compromised controllers across roughly 700 European supermarkets. The warning had named the flaws, but only each operator could know which refrigerator was still listening.