NIST Bets on AI to Clear AI-Created CVE Backlog

NIST is looking at deploying artificial intelligence to help process vulnerabilities at the National Vulnerability Database — the same category of AI-powered tooling that is flooding the pipeline with more CVEs than it can enrich on a human timescale.

Dark Reading reported Thursday that vulnerability volumes are surging, driven by AI-augmented scanning and research tools that find bugs faster than disclosure pipelines were designed to handle. NIST is now asking whether AI can be the answer to its own backlog.

This isn’t a new problem. It is a sharper version of one. In 2023 and 2024, NVD fell into a prolonged enrichment backlog — the step in the pipeline where a CVE ID receives CVSS scores, Common Weakness Enumeration tags, and product references so that vulnerability management tools and patching workflows can actually act on it. CISA stepped in as a processing partner to clear part of the queue. The backlog shrank, then AI-assisted research tooling got meaningfully better, and the intake volume is climbing again.

The structural problem here is that AI-augmented vulnerability research scales with compute in a way human advisory writing does not. A researcher with a well-configured fuzzer and an LLM for triage and writeup can disclose in a weekend what used to take months of manual analysis. We saw a compressed version of this in August when a single researcher dropped ten CVEs across MCP server implementations in a single day — each one real, each one requiring NVD review, most across projects where maintainers weren’t expecting coordinated disclosure. That pattern, replicated at scale across the open source ecosystem, is what NIST is staring at.

The downstream effect that matters most is CVSS. Security teams have built their patch prioritization workflows around NVD enrichment data, and primarily around CVSS scores. An unenriched CVE sits in the database with a blank or placeholder score. A vulnerability management platform pulling from NVD returns incomplete signal. In a world where patch capacity is finite and the queue is measured in thousands — August’s Patch Tuesday alone logged 88 CVEs from Microsoft — a blank CVSS score gets deprioritized by default and may never get worked.

NIST’s proposed direction — using AI to automate enrichment through auto-scoring, CWE tagging, and product matching — is a reasonable structural response to a structural problem. It is also one that introduces its own accuracy questions. Assigning a CVSS score is contextual work: it requires considering what exploitation actually requires, whether authentication preconditions exist, how the affected software is typically deployed, and what the realistic attacker population looks like. Whether current models can do that consistently enough to be trusted as a primary enrichment source is genuinely open, and NIST hasn’t answered it.

Getting enrichment wrong in the direction of “too high” creates alert fatigue in SOCs already operating near capacity. Getting it wrong “too low” buries real exposure. Both failure modes are worse than a blank field in different ways, and the industry’s ability to detect a quiet accuracy degradation — as opposed to a spectacular enrichment failure — is not well developed.

Analysis: The interesting part of this is not the irony of AI creating a problem and AI being asked to solve it; that’s been the pattern in security tooling for as long as there have been security tools. The interesting part is structural: the entire patch-prioritization apparatus the industry has built assumes NVD enrichment is reliable and reasonably timely. Both assumptions are under simultaneous stress. NIST turning to AI doesn’t resolve the underlying mismatch — it relocates where the errors happen.

Defenders layering EPSS scores alongside CVSS, or using CISA KEV additions as a primary triage signal, are somewhat insulated from this. The KEV catalog is maintained reactively and at lower volume, with human review before anything gets added. But for vulnerabilities that never reach KEV — the large majority of disclosed CVEs — NVD enrichment quality matters more than the current moment makes obvious.

This is worth watching not because AI-assisted CVE enrichment will fail visibly, but because it may degrade in ways that are quiet and slow and only apparent well after they’ve distorted someone’s risk posture.