AI Slop Is Breaking Vulnerability Management, and Open Source Is Paying the Price
Cybersecurity has a new noise problem, and this time it is dressed up as “AI-powered research.”
Open source maintainers are being flooded with AI-generated vulnerability reports that look convincing, use the right security words, and sometimes even include CVSS-style impact claims. The problem? Many of them are nonsense. They are not reproducible, not exploitable, and sometimes not even technically coherent.
That might sound harmless until you remember who has to review them: real maintainers, often unpaid, already stretched thin, and responsible for software that half the internet depends on.
Daniel Stenberg , maintainer of curl, called this trend “death by a thousand slops,” describing how AI-generated vulnerability reports waste security team time while rarely producing real findings. Seth Larson also warned about the rise of low-quality, hallucinated security reports hitting major open source projects. These reports may be fake, but the time wasted triaging them is very real. (daniel.haxx.se). And now the impact is spreading beyond maintainers.
On April 15, 2026, NIST changed how the National Vulnerability Database handles CVE enrichment. To be clear, NVD is not refusing CVEs. But it will no longer immediately enrich every CVE with full metadata. Instead, it will prioritize vulnerabilities in CISA KEV, federal government software, and critical software. Lower-priority CVEs may still be listed, but without immediate enrichment such as CVSS, CPE, CWE, and affected product mapping. (NIST)
That is a big deal.
Most vulnerability management tools, open source scanners, dashboards, CI/CD gates, and Jira automation workflows depend on enriched NVD data. When that enrichment is delayed or missing, security teams lose context. Severity may be unclear. Affected versions may be missing. Product mappings may be incomplete. Suddenly, teams are left asking: is this a real risk, a low-priority issue, or just another garbage report that slipped into the ecosystem?
This is where AI slop becomes dangerous. It does not just annoy maintainers. It pollutes the signal chain that security teams depend on.
Projects are already reacting. Node.js introduced a HackerOne Signal requirement to reduce low-quality vulnerability submissions. OpenSSF is actively discussing best practices for handling AI-generated security reports, including stronger accountability and quality expectations. (Node.js)
Honestly, this should disappoint all of us.
AI could have been a force multiplier for security research. It can help analyze code, summarize advisories, generate test cases, and assist skilled researchers. But instead, some people are using it as a vulnerability vending machine: generate report, submit, hope for bounty, repeat. That is not research. That is spam with a security label.
The way forward is not to ban AI. The way forward is to raise the bar.
A valid vulnerability report should include clear reproduction steps, affected versions, technical impact, exploitability conditions, and evidence that the reporter understands the issue. Bug bounty platforms should penalize repeated low-quality AI-generated submissions. Maintainers should be allowed to require proof-of-concept evidence. Security teams should stop treating CVE IDs as complete truth and start using multiple signals: CISA KEV, EPSS, vendor advisories, OSV, GitHub Advisory Database, exploit availability, asset exposure, and runtime context.
The uncomfortable truth is this: CVE is an identifier, not a risk decision. NVD is an enrichment source, not an oracle.
For DevSecOps teams, this means updating pipelines and dashboards now. Do not block only on CVSS. Do not ignore unscored vulnerabilities. Do not assume NVD enrichment will always arrive quickly. Build a fallback process for “unknown but potentially serious” vulnerabilities.
AI did not break cybersecurity by itself. Lazy, unverified, mass-produced AI submissions did.
Open source maintainers deserved better. NVD analysts deserved better. Security teams deserved better. If we want AI to help cybersecurity, we need to stop rewarding people for generating noise and start rewarding people who prove real risk because cybersecurity already has enough real problems.We do not need fake ones manufactured at scale.
