# CVSS says how bad a flaw is and EPSS says how likely it is to be attacked; rank by likelihood first and use severity to set the floor

COMPARISON, read 10 Nov 2026

**The short answer.** CVSS scores severity from 0.0 to 10.0 with a published formula; EPSS gives a daily probability that a CVE will be exploited in the next 30 days. In the KEV Lag Ledger, all 63 listings with EPSS of 0.1 or more also scored CVSS 7.0 or above, while 173 listings scored 7.0 or above with EPSS under 0.1.

- **Maintained by:** CVSS: FIRST CVSS-SIG. EPSS: FIRST EPSS SIG, scores generated by Empirical Security
- **Scale:** CVSS: 0.0 to 10.0. EPSS: probability 0 to 1 plus percentile
- **Current version:** CVSS v3.1 and v4.0. EPSS model v2026.06.15 (CSV header, 2026-11-10)
- **Updated:** CVSS: when a CNA or NVD scores or rescores. EPSS: daily
- **Answers:** CVSS: the severity of the flaw. EPSS: the chance of exploitation in the next 30 days

## The two questions

CVSS and EPSS answer different questions, and the specifications say so. The CVSS specification describes its outputs as scores indicating the severity of a vulnerability relative to other vulnerabilities, with a Base Score that assumes the reasonable worst case impact. FIRST's EPSS page describes a model that estimates the probability that a published CVE will be exploited in the wild in the next 30 days.

The two are not rivals for the same job. A CVSS score is fixed by the flaw's characteristics and moves only when someone rescores. An EPSS score is recalculated daily from about 2,850 features and moves as exploit code, discussion and other signals appear. One is a property of the flaw; the other is an estimate about attackers.

| Property | CVSS v3.1 / v4.0 | EPSS |
| --- | --- | --- |
| Output | Severity score 0.0 to 10.0 and a band (Low to Critical) | Probability 0 to 1 and a percentile |
| Method | Published formula (v3.1) or MacroVector lookup (v4.0) | Machine-learning model; training data and pipeline not published |
| Refresh | On rescoring | Daily, after 13:30 UTC |
| Organization-specific | Environmental metrics, optional | Population-level; FIRST says it is not tailored to an organization |

## What each gets right

CVSS is reproducible. With the Table 16 constants any analyst can recompute a v3.1 Base Score: ISS = 1 - (1 - C)(1 - I)(1 - A), Impact = 6.42 x ISS for unchanged scope, Exploitability = 8.22 x AV x AC x PR x UI, and BaseScore = Roundup(Min[Impact + Exploitability, 10]). An auditor can check the arithmetic. Scores also exist for nearly every CVE within days of publication.

EPSS is calibrated. FIRST states that a score of 0.05 means a 5 percent probability, about 1 in 20 within 30 days across the population, and that low scores cluster most vulnerabilities because exploitation is rare. The base rate of exploitation is roughly 2 to 3 percent, so a probability can be compared with a base rate. A CVSS 7.0 cannot be compared with any rate.

Each has a documented weak point. FIRST lists the EPSS limits itself: scores are population-level, a newly disclosed vulnerability has a sparse record, and the training labels reflect only the exploitation that data partners can observe. For CVSS the specification states that the Base Score assumes the reasonable worst case, and that the qualitative bands are optional. Neither weakness is hidden, and both can be written into a policy as stated assumptions.

## What the ledger shows

The KEV Lag Ledger holds 318 vulnerabilities that CISA confirmed as exploited between 2025-11-10 and 2026-11-09. Every one has a CVSS score: median 8.8, with 285 (89.6 percent) at 7.0 or above and 33 (10.4 percent) below. Of the 318, 260 have an EPSS value from the listing day: median 0.0066, with 197 (75.8 percent) below 0.1.

Cross the two for those 260. 63 listings had both CVSS 7.0 or above and EPSS 0.1 or above. 173 had CVSS 7.0 or above and EPSS below 0.1. 24 had neither. None had EPSS at 0.1 or above with CVSS below 7.0. At a CVSS cutoff of 9.0, the counts are 35 (both), 88 (CVSS only), 28 (EPSS only) and 109 (neither). A CVSS High cutoff captured more of these confirmed exploited vulnerabilities than an EPSS 0.1 cutoff did.

| Listing-day cell (260 KEV additions with EPSS) | CVSS 7.0 or above | CVSS below 7.0 |
| --- | --- | --- |
| EPSS 0.1 or above | 63 | 0 |
| EPSS below 0.1 | 173 | 24 |

## What the ledger cannot show

This is coverage, not efficiency. The ledger holds only vulnerabilities that were exploited, so it can say how many a threshold would have caught but not how many unexploited vulnerabilities the threshold would also have queued. A CVSS cutoff of 7.0 catches 89.6 percent of the 318 and also queues whatever share of all published CVEs scores 7.0 or above, a denominator the ledger does not hold.

FIRST's How It Works page frames this as the coverage and efficiency tradeoff, with CVSS-based thresholds plotted as points on the same axes as the EPSS curve, and states that the EPSS curve sits substantially above both the random baseline and the CVSS points at most coverage levels. That is FIRST's published finding, and its plots should be read for the threshold you intend to use.

## How to combine them

FIRST's Using EPSS page warns against multiplying EPSS by an ordinal such as CVSS: EPSS is a calibrated probability, CVSS is an ordinal ranking with no empirical calibration, and the product has no interpretable meaning. Combine them by rule, not by arithmetic.

A workable rule set has three lines. Rank by EPSS probability, with a documented threshold, for everything not already confirmed as exploited. Use CVSS to set the minimum service level by asset tier. Keep CVSS Critical as an escalation when no EPSS value exists yet, since 58 of the 318 ledger additions had none on their listing day.

The ledger lets a threshold be sized. At an EPSS cutoff of 0.01, 116 of the 260 scored listings (44.6 percent) were flagged. At the 90th percentile or higher, 75 (28.8 percent). At 0.1, 63 (24.2 percent). At 0.5, 34 (13.1 percent). A lower cutoff catches more of the confirmed set and sends more of the whole CVE population to the queue, and only the second half of that tradeoff depends on your environment, so measure it on your own open findings.

## Reading the result

The ledger supports one clear statement for a policy: neither number alone flagged the exploited set on listing day. EPSS at 0.1 or above flagged 24.2 percent of the 260 scored listings, CVSS at 9.0 or above flagged 45.9 percent of the 318, and CVSS at 7.0 or above flagged 89.6 percent.

See the CVSS against KEV and EPSS against KEV pages for those comparisons, and the three-way page for the ordering the ledger supports.

For an audit file, record the EPSS model version from the first line of the CSV (v2026.06.15 on 2026-11-10), the date of the score, and the CVSS version with its vector. FIRST notes that the model changes about annually, so a score from before a retraining is not comparable with one after it.

## Verdict by situation

- **Federal agencies and contractors:** Rank by KEV status first, since BOD 26-04 builds its timelines on it, and use CVSS technical severity and EPSS only for the CVEs not in the catalog.
- **Large enterprises with ten or more scanners:** Rank by EPSS probability first with a documented threshold, because CVSS Critical alone would queue far more work than a large enterprise can close.
- **Managed security providers:** Rank by EPSS first across clients, because it is one public, daily, comparable number, and report CVSS beside it because every client contract already quotes it.
- **Mid-size teams replacing Kenna or a spreadsheet:** Rank by EPSS first to replace the exploitation-likelihood part of the old risk score, and keep CVSS as the severity floor the old score hid.
- **Application security and infrastructure in one queue:** Rank infrastructure CVEs by EPSS first and application findings by CVSS first, because EPSS covers published CVEs and most application findings have none.

## Other comparisons

- [CVSS vs KEV](/compare/cvss-vs-kev/)
- [EPSS vs KEV](/compare/epss-vs-kev/)
- [SSVC vs CVSS](/compare/ssvc-vs-cvss/)
- [SSVC vs EPSS](/compare/ssvc-vs-epss/)
- [CVSS v3.1 vs v4.0](/compare/cvss-v3-1-vs-v4-0/)
- [KEV vs vendor intelligence](/compare/kev-vs-vendor-intelligence/)
- [CVSS vs EPSS vs KEV](/compare/cvss-vs-epss-vs-kev/)

## Sources

- FIRST, CVSS v3.1 specification (section 7.1, Table 16): https://www.first.org/cvss/v3.1/specification-document (read 10 Nov 2026)
- FIRST, CVSS v4.0 specification: https://www.first.org/cvss/v4-0/specification-document (read 10 Nov 2026)
- FIRST, EPSS home: https://www.first.org/epss/ (read 10 Nov 2026)
- FIRST, How EPSS works (coverage and efficiency, limitations): https://www.first.org/epss/how-it-works (read 10 Nov 2026)
- FIRST, Using EPSS (score laundering, base rate): https://www.first.org/epss/using-epss (read 10 Nov 2026)
- EPSS daily CSV (model version header, 2026-11-10): https://epss.cyentia.com/epss_scores-current.csv.gz (read 10 Nov 2026)

Last reviewed 10 Nov 2026. Canonical: https://epssvscvss.com/compare/epss-vs-cvss/
