NKO Reality Check #002 · Vulnerability Severity Disagreement
We compared official severity scores for 2,000 real vulnerabilities. Where more than one source scored the same CVE, they disagreed on severity category 53% of the time — including 16 cases where the gap meant the difference between "fix eventually" and "fix now".
Every CVE can carry more than one official CVSS severity score — NVD's own analysts, the vendor's own PSIRT, GHSA (GitHub), and Authorized Data Publishers like CISA-ADP can each publish their own. We pulled 2,000 real CVEs published between 2026-07-15 and 2026-08-30 straight from NVD's own API. Only 188 (9.4%) had more than one independent score to compare — of those, 100 (53.2%) landed in a different severity category depending on which source you read, and 16 jumped two full categories at once (e.g. "Medium" on one side, "Critical" on the other).
Check your own CVE — free, no account
Enter a CVE ID. We show every severity score we can find, its source, and whether they agree.
What we found, in numbers
Method
Ground truth for "how many independent scores exist": NVD's own
API 2.0 record for each CVE, which already lists every CVSS
score it has received, tagged with the real organization that
submitted it (nvd@nist.gov, vendor PSIRTs,
security-advisories@github.com for GHSA, CISA-ADP,
etc.) — cross-referenced by CVE ID, never by text matching.
CVSS v2 was excluded (0 of the 188 cases used it) because its
severity bands aren't directly comparable to v3.x/v4.0; v3.0,
v3.1 and v4.0 share the same official band cutoffs (Low
0.1-3.9, Medium 4.0-6.9, High 7.0-8.9, Critical 9.0-10.0), so
comparing across those versions is valid.
We also checked whether the gap has a practical consequence beyond the category label: for 8 of these cases, we compared the full advisory — CVSS, CWE, affected-product scope — between GHSA and NVD field by field. In most cases the underlying data (weakness type, affected versions) was identical; the disagreement is specifically about how severe the same technical facts are judged to be, not about missing information. Full write-up, all 188 cases, and the 8-case field comparison: github.com/NKO-LLC-TECH/nko.
16 real cases where the category itself changes — plus contrast
16 real two-category jumps (highlighted), plus 2 one-category jumps and 2 cases of full agreement for contrast — every row links to the real NVD record so anyone can verify it independently.
| CVE | Scores by source | Categories seen | Gap | Evidence |
|---|
Batch checks
Free: up to 50 CVE IDs per request, no account, same engine as the single check above. Because NVD's own public API is rate-limited, batches are queued and processed in the background — you get a signed link to poll for the result. Paid: up to 10,000 CVEs for €2,500, delivered within 24-48h — every CVE's full source-by-source breakdown, agreement status, and category gap, ready to cross-reference against your own portfolio.
€2,500 — batch up to 10,000 CVEs
For checking a whole vulnerability portfolio at once — built for security-rating platforms, vulnerability management tools, and cyber-insurance underwriting, not single lookups. Same textarea above works for this — paste your list, confirm, pay, get a signed link to your results within 24-48h. No "contact sales" — this is the price.