CVE JSON CVSS Metrics: cvssV4_0, cvssV3_1, baseScore Trap

How the metrics array in a CVE JSON 5 record actually looks across all 96,879 CVEs published in 2026: which CVSS versions CNAs ship, when CISA-ADP fills the gap, what scenarios mean, why cvssV4_0.baseScore is sometimes a Base+Threat score, and a tested script that picks one score per record and says where it came from.

Black-and-white close-up of two round pressure gauges with needles and a regulator valve on old industrial pipework
Photo by Miguel Ángel Padriñán Alba on Unsplash

Short answer. A CVE JSON 5 record can hold zero, one or several CVSS scores, from the CNA and from Authorized Data Publishers such as CISA-ADP, in different CVSS versions and for different scenarios. Across the 96,879 CVEs published in 2026, under a third carry a CNA CVSS v4.0 score, a fifth carry no CNA score at all, and the field called baseScore in a v4.0 entry sometimes holds a Base+Threat score. If your pipeline reads “the CVSS score” from a record, it is making several choices whether you wrote them down or not. This article shows what the field-level data looks like and gives a tested selection script.

Scope. We read CVE_Record_Format.json and imports/cvss/cvss-v4.0.json from the CVEProject/cve-schema main branch on November 13, 2026, and streamed every record in the cvelistV5 release cve_2026-11-13_0000Z (2026-11-13_all_CVEs_at_midnight.zip). Counts cover records in state PUBLISHED with a datePublished in 2026. Scores were recomputed with the Red Hat cvss Python library v3.6 on Python 3. For how to read a v4.0 vector itself, see How to Read a CVSS v4.0 Vector String; for where these scores sit in a patch decision, see the vulnerability prioritization workflow.

What the schema says

metrics lives in a container: the CNA’s (containers.cna.metrics) or an ADP’s (containers.adp[].metrics). It is not required in the CNA published container, whose required fields are providerMetadata, descriptions, affected and references.

When present, metrics is an array (minItems: 1). The schema’s description of an entry is “impact type information (e.g. a text description, CVSSv2, CVSSv3, CVSSV4, etc.)”, and it uses anyOf to require at least one of these keys:

KeyContentValidated against
cvssV4_0CVSS v4.0 objectimports/cvss/cvss-v4.0.json (FIRST schema)
cvssV3_1CVSS v3.1 objectimports/cvss/cvss-v3.1.json
cvssV3_0CVSS v3.0 objectimports/cvss/cvss-v3.0.json
cvssV2_0CVSS v2.0 objectimports/cvss/cvss-v2.0.json
other{type, content}: “A non-standard impact description, may be prose or JSON block”free-form

Two optional siblings matter for parsing:

  • format: “Name of the scoring format.” In practice, 55,806 CNA entries in 2026 set it to CVSS, so do not use it to detect the version. Use the key name.
  • scenarios: an array of {lang, value}. The schema says: “If no specific scenario is given, GENERAL is used as the default and applies when no more specific metric matches.”

The v4.0 object requires version, vectorString, baseScore and baseSeverity, and allows the individual metric fields: Base, Threat (exploitMaturity), Environmental (modified* and the requirement fields) and Supplemental (Safety, Automatable, Recovery, valueDensity, vulnerabilityResponseEffort, providerUrgency). There is no threatScore or environmentalScore field. That gap causes the trap described below.

ADP containers use the same metrics structure. CISA-ADP also uses other for its own data: in 2026 records we counted 90,133 other entries of type ssvc and 215 of type kev. Those are what CISA Vulnrichment’s SSVC values and the KEV date come from, so a parser that keeps only CVSS keys throws them away.

What 2026 records actually contain

CNA CVSS formats present (2026 published records)RecordsShare
v3.1 only45,50747.0%
none20,40821.1%
v3.1 + v4.012,45812.9%
v4.0 only11,29611.7%
v2.0 + v3.0 + v3.1 + v4.06,3706.6%
v3.0 only6120.6%
v3.0 + v4.02160.2%
v2.0 + v3.1 + v4.012<0.1%
Total96,879

Totals across those rows: 30,352 records (31.3%) had a CNA v4.0 entry, and 64,347 (66.4%) a CNA v3.1 entry.

v4.0 adoption is concentrated. VulnCheck (10,176 records), VulDB (6,370) and GitHub (4,287) published 68.6% of the CNA v4.0 records. VulDB’s 6,370 v4.0 records match the 6,370 records in the four-version row, and the VulDB record we opened (CVE-2026-82598) lists v4.0, v3.1, v3.0 and v2.0 entries side by side. A parser that takes the first entry in the array gets whichever one the CNA happened to list first.

Of the 20,408 records with no CNA CVSS, 15,148 had a CVSS v3.1 entry from CISA-ADP, and 5,260 had no CVSS from either the CNA or CISA-ADP. (A few of those carry a supplier ADP score or a CNA other rating such as “Red Hat severity rating”, which appeared in 1,476 CNA entries.) Only 4 CISA-ADP entries in 2026 used v4.0.

The baseScore trap

The CVSS v4.0 specification names scores by the metric groups used: CVSS-B (Base), CVSS-BT (Base and Threat), CVSS-BE and CVSS-BTE. Our vector-string article covers that nomenclature. The CVE schema, though, gives the v4.0 object only one score field, baseScore. When a CNA puts a Threat metric in the vector, it has to decide which number goes in that field. CNAs decide differently.

We took every CNA cvssV4_0 entry from 2026 whose vectorString sets Exploit Maturity to something other than X, recomputed the score with and without the E: metric, and compared both to the stored baseScore:

Stored baseScore equals…Entries
Base-only score (CVSS-B)5,816
Base+Threat score (CVSS-BT)1,067
Both equal (E metric did not change the score)68
Neither3
Total with E set6,954

The split follows the CNA. VulDB accounts for 5,790 of the Base-only entries. GitHub (432), Palo Alto Networks (120), QNAP (106) and NETGEAR (48) lead the CVSS-BT side. Two real records show the difference:

  • CVE-2026-82598 (VulDB): vector CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:L/VI:L/VA:L/SC:N/SI:N/SA:N/E:P, stored baseScore 6.9. The full vector computes to 5.5; without E:P, 6.9. The stored number is CVSS-B.
  • CVE-2026-41129 (GitHub, a Craft CMS SSRF): vector CVSS:4.0/AV:N/AC:L/AT:N/PR:H/UI:N/VC:H/VI:L/VA:N/SC:N/SI:N/SA:N/E:P, stored baseScore 5.5, severity MEDIUM. Without E:P the score is 7.0 (HIGH). The stored number is CVSS-BT.

Both are valid by the schema. But if your SLA says “High and above within 14 days”, the second record misses the bar because the CNA folded in Exploit Maturity, and the first meets the bar or misses it purely on base impact. If you also add an exploitation signal from KEV or EPSS, a CVSS-BT number counts exploit evidence twice in one direction or the other. Decide which you want, recompute it from the vector, and label it.

Exploit Maturity values seen in CNA v4.0 entries in 2026: NOT_DEFINED 10,620, field absent 19,051, UNREPORTED 482, PROOF_OF_CONCEPT 210 and ATTACKED 41. Those counts come from the exploitMaturity JSON field, which not every producer fills even when the vector sets E:. Read the vector, not the field.

Scenarios: one CVE, several scores

3,543 CNA metrics entries in 2026 named a scenario other than GENERAL. F5’s CVE-2026-41225 is a clean example. It has three entries:

ScenarioFormatScoreVector
GENERALcvssV3_17.2CVSS:3.1/AV:N/AC:L/PR:H/UI:N/S:U/C:H/I:H/A:H
GENERALcvssV4_08.6CVSS:4.0/AV:N/AC:L/AT:N/PR:H/UI:N/VC:H/VI:H/VA:H/SC:N/SI:N/SA:N
Appliance ModecvssV3_19.1CVSS:3.1/AV:N/AC:L/PR:H/UI:N/S:C/C:H/I:H/A:H

If you run the product in the named configuration, the scenario-specific score is the one that applies to you. A dashboard that shows the maximum overstates risk for everyone else, and one that shows the first entry may understate it for appliance-mode customers. Pick GENERAL by default, and let asset metadata select a named scenario when it matches.

CNA and CISA-ADP disagree more than you would expect

CISA-ADP mostly adds CVSS where the CNA has none (see ADP containers). In 468 records from 2026 both published a v3.1 score. In 58 of those (12.4%) the base scores differed: CISA-ADP was higher in 31 and lower in 27. Some gaps are large. In CVE-2026-41080 (libexpat hash flooding, assigned by MITRE), the CNA entry is AV:L/AC:H/PR:N/UI:N/S:U/C:N/I:N/A:L (2.9) and CISA-ADP’s is AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:N/A:N (7.5). Those two analysts read the impact very differently. Averaging them would produce a number nobody asserted. Store both, with provider, and let your policy say which one drives the SLA.

A tested script: pick one score and keep its provenance

This script applies a fixed order: CNA before CISA-ADP, v4.0 before v3.1 before v3.0, GENERAL (or unlabelled) scenarios before named ones. For v4.0 it recomputes CVSS-B and CVSS-BT from the vector and says which one baseScore holds. It needs pip install cvss (we used v3.6).

"""Pick one CVSS entry per CVE JSON 5.x record, and say where it came from.
Order: CNA v4.0 > CNA v3.1 > CNA v3.0 > CISA-ADP v3.1. GENERAL scenario first.
Requires: pip install cvss (Red Hat library, v3.6 tested)."""
import json, sys
from cvss import CVSS4

ORDER = ["cvssV4_0", "cvssV3_1", "cvssV3_0"]

def is_general(entry):
    sc = entry.get("scenarios")
    return not sc or any(s.get("value", "").upper() == "GENERAL" for s in sc)

def candidates(record):
    cna = record["containers"]["cna"].get("metrics", [])
    for fmt in ORDER:
        for e in sorted(cna, key=lambda e: not is_general(e)):
            if fmt in e:
                yield "cna", fmt, e
    for adp in record["containers"].get("adp", []):
        if adp["providerMetadata"].get("shortName") == "CISA-ADP":
            for e in adp.get("metrics", []):
                if "cvssV3_1" in e:
                    yield "CISA-ADP", "cvssV3_1", e

def pick(record):
    for source, fmt, e in candidates(record):
        m = e[fmt]
        out = {"source": source, "format": fmt, "stored": m["baseScore"],
               "vector": m["vectorString"],
               "scenario": [s["value"] for s in e.get("scenarios", [])] or ["(none)"]}
        if fmt == "cvssV4_0":
            parts = m["vectorString"].split("/")
            base_only = "/".join(p for p in parts if not p.startswith("E:"))
            out["cvss_b"] = float(CVSS4(base_only).base_score)
            out["cvss_bt"] = float(CVSS4(m["vectorString"]).base_score)
            out["stored_is"] = ("B and BT equal" if out["cvss_b"] == out["cvss_bt"] else
                                "CVSS-B" if out["stored"] == out["cvss_b"] else
                                "CVSS-BT" if out["stored"] == out["cvss_bt"] else "neither")
        return out
    return {"source": None, "note": "no CVSS from CNA or CISA-ADP"}

if __name__ == "__main__":
    for path in sys.argv[1:]:
        rec = json.load(open(path))
        print(rec["cveMetadata"]["cveId"], json.dumps(pick(rec)))

We ran it on record files fetched from raw.githubusercontent.com/CVEProject/cvelistV5/main/cves/2026/... on November 13, 2026 (output trimmed to the key fields):

CVE-2026-1995   source=CISA-ADP format=cvssV3_1 stored=7.8
CVE-2026-41080  source=cna format=cvssV3_1 stored=2.9 scenario=GENERAL
CVE-2026-41129  source=cna format=cvssV4_0 stored=5.5 cvss_b=7.0 cvss_bt=5.5 stored_is=CVSS-BT
CVE-2026-41225  source=cna format=cvssV4_0 stored=8.6 cvss_b=8.6 cvss_bt=8.6 stored_is="B and BT equal" scenario=GENERAL
CVE-2026-82598  source=cna format=cvssV4_0 stored=6.9 cvss_b=6.9 cvss_bt=5.5 stored_is=CVSS-B
CVE-2026-86471  source=cna format=cvssV4_0 stored=9.3 cvss_b=9.3 cvss_bt=9.3 stored_is="B and BT equal" scenario=GENERAL

CVE-2026-1995 (an IDrive Cloud Backup Client privilege escalation, CNA CERT/CC) has no CNA CVSS, so the CISA-ADP v3.1 score is used and labelled. CVE-2026-86471 (Apache CouchDB) sets E:A; Attacked is the worst-case Exploit Maturity, so Base and Base+Threat are equal.

This is a starting point, not a policy. You might prefer v3.1 over v4.0 while your SLAs are still written in v3.1 bands (the two versions’ scores are not comparable), keep the CISA-ADP score alongside the CNA one, or pick a named scenario from asset data.

Common mistakes

  • Reading metrics[0]. The array order is the CNA’s choice; some records list four CVSS versions.
  • Detecting the version from format. Most entries just say CVSS. Use the key (cvssV4_0, cvssV3_1, …) or the version field.
  • Treating v4.0 baseScore as CVSS-B. Recompute from vectorString if your policy cares.
  • Dropping other. CISA-ADP’s SSVC and KEV entries are other; so are vendor ratings such as Red Hat’s.
  • Overwriting CNA with ADP (or the reverse). Keep both with providerMetadata.shortName.
  • Ignoring scenarios. The highest score may belong to a configuration you do not run.

Limits of this article

Counts are a snapshot of cvelistV5 at 00:00 UTC on November 13, 2026, restricted to records published in 2026. CNAs update records, so numbers will drift. Score recomputation used one library; where we compared stored and recomputed values, three entries matched neither, and we did not investigate them. We do not say which CNA practice is correct. The schema permits both, and the CVSS v4.0 specification’s naming guidance applies to how scores are displayed, not to this JSON field.

Sources and verification

Important claimSourceStatus
metrics optional in CNA container; array structure; anyOf of five keys; format, scenarios and GENERAL default; other = {type, content}CVE_Record_Format.json, CVEProject/cve-schema (read Nov 13, 2026)Verified
v4.0 object required fields; no Threat/Environmental score fields; Supplemental field namescvss-v4.0.json import, CVEProject/cve-schemaVerified
CVSS-B / BT / BE / BTE nomenclatureCVSS v4.0 Specification Document, FIRSTVerified in our Oct 17, 2026 article; not re-read today
All 2026 counts (formats, CISA-ADP fill, other types, scenarios, B vs BT split, CNA vs ADP disagreement, CNA concentration)Our script over cvelistV5 release cve_2026-11-13_0000ZQualified: our computation, snapshot date stated
Example records CVE-2026-1995, -41080, -41129, -41225, -82598, -86471 (vectors, stored scores, scenarios)cvelistV5 record filesVerified
Recomputed CVSS-B / CVSS-BT valuesRed Hat cvss library v3.6, run Nov 13, 2026Qualified: one implementation

Frequently asked questions

Is the metrics field required in a CVE record?

No. In CVE_Record_Format.json the CNA published container requires providerMetadata, descriptions, affected and references; metrics is optional. When it is present it is an array with at least one entry, and each entry must contain at least one of cvssV4_0, cvssV3_1, cvssV3_0, cvssV2_0 or other. In 2026, 21.1% of published records had no CNA CVSS entry.

Does cvssV4_0.baseScore always contain the CVSS-B score?

No. The schema has only baseScore for v4.0, so producers that set Threat metrics such as E:P make a choice. In the 2026 corpus, 5,816 CNA v4.0 entries with Exploit Maturity set stored the Base-only score and 1,067 stored the Base+Threat score. Recompute from vectorString if the distinction matters to your policy.

What does the scenarios field mean in CVE metrics?

It describes the situation a metrics entry applies to. The schema says that if no specific scenario is given, GENERAL is the default and applies when no more specific metric matches. In 2026 records, 3,543 CNA metrics entries named a specific scenario, for example F5's Appliance Mode score for CVE-2026-41225.

Should I use the CNA score or the CISA-ADP score?

Use the CNA score when there is one and keep the CISA-ADP score as a second, labelled value. CISA-ADP mostly fills gaps: it supplied v3.1 for 15,148 of the 20,408 2026 records with no CNA CVSS. Where both existed, 58 of 468 records had different v3.1 base scores, sometimes far apart.

Vulnerability ManagementCveCve Json 5CvssCvss V4Cisa VulnrichmentAdpParsing