CVSS 3.0 and 3.1 agree on every base score; a floating-point ceiling does not

All 259,200 CVSS base and temporal vectors score the same under 3.0 and 3.1. Environmental scores and naive 3.0 rounding are where they part.

Security teams hold vulnerability advisories scored under CVSS 3.0 and CVSS 3.1, often side by side in one scanner report. Two questions come up again and again. Can a 3.0 score be compared with a 3.1 score? And why do two advisories for the same bug carry different numbers? We built cvss-vector-check to answer both from the vectors themselves. Then we used it to score the whole formula space.

The measurement

A CVSS v3 base vector has eight metrics, giving 2,592 combinations. Adding the three temporal metrics gives 259,200. We scored every one of them under the v3.0 and v3.1 formulas. We also scored 200,000 random vectors with every environmental metric set. Each score was checked against Red Hat's open-source cvss Python package (version 3.6), which is a separate implementation. It agreed on every score in both versions. The scripts are in the venture's oracle/ folder, and 6,000 of the oracle's scores are committed as a test fixture that runs on every change.

QuestionResult
Base score differs between 3.0 and 3.10 of 2,592
Temporal score differs between 3.0 and 3.10 of 259,200
Environmental score differs12,623 of 200,000 (6.3%), all with a Changed modified scope
Size of that differencealways 0.1: 11,155 higher under 3.1, 1,468 lower
Lands in a different severity band610 of 200,000 (0.3%)
Naive Math.ceil(x*10)/10 3.0 rounding wrong228 of 259,200 temporal scores, all +0.1; 0 base scores

The random sample picks every metric value with equal odds. Real advisories are not spread like that, so 6.3% is the share of the formula space, not of published scores.

What it means

Base scores are interchangeable. If both are computed correctly, a CVSS:3.0 vector and the same metrics labelled CVSS:3.1 give the same base score. So do their temporal scores. A different base score between two advisories means they chose different metrics, not different versions.

Environmental scores move only with a Changed modified scope. CVSS 3.1 rewrote one formula, the Modified Impact for a Changed scope. Everywhere else the two versions match. Where the formula does change, the shift is 0.1 either way. That is small, but it moved one vector in 330 to a new severity band. One example from the sample scores 7.0 (High) under 3.1 and 6.9 (Medium) under 3.0.

A 3.0 calculator can be 0.1 high for a dull reason. CVSS 3.0 says "round up to one decimal" in words. In floating point, 5.0 × 0.92 is stored as 4.6000000000000005, so Math.ceil(x * 10) / 10 gives 4.7. The oracle, and this tool, give 4.6, which is what the arithmetic means. CVSS 3.1 redefined Roundup with integer arithmetic to remove exactly this. The naive form is wrong on 228 temporal scores and on no base score, and none of those errors changes the severity band.

Our own first draft of the engine used the naive ceiling for 3.0. The oracle fixture caught it: 7 of the first 3,000 random 3.0 vectors disagreed. The fix reads the 3.0 rule mathematically. The tool keeps the naive variant on hand, so when an audited report states 4.7 for a vector worth 4.6, it can say where the 4.7 came from.

The tool

cvss-vector-check runs in the browser and as a zero-dependency CLI. It does four things:

  • It scores and explains one vector.
  • It compares two vectors metric by metric and shows what each difference does on its own.
  • It shows how the score would change if one metric changed.
  • It audits a pasted report, CSV, or NVD or OSV JSON. Every vector is rescored and checked against the score and severity label written next to it. When a stated score is really the temporal score, the other version's score or the naive-rounding score, the tool says so.

node cli.mjs --audit report.json --strict exits 1 on a mismatch, so it can run as a CI step over generated advisories.

While testing, a repository lint (inherited_keys.py) flagged a lookup in the parser. A vector containing AV:toString parsed as valid because a plain object's in check also sees Object.prototype. It now refuses it, and a test covers it.

Next, if people use it: CVSS 4.0, which scores through FIRST's macrovector lookup table rather than a formula, checked the same way against an independent implementation.

Produced with AI assistance. Not affiliated with FIRST, NIST, the NVD, MITRE or Red Hat. A CVSS score measures technical severity, not risk.

All writing · More in Tools and projects · All tools and research