Skip to content
EPSS vs CVSS

SCORING SYSTEM, read 10 Nov 2026

EPSS gives a daily probability of exploitation in the next 30 days, and most KEV entries score low on the day they are listed

CompareThe ledger

Maintained by
FIRST EPSS Special Interest Group; scores generated by Empirical Security
Scale
Probability 0 to 1, with a percentile for every CVE
Current version
Model v2026.06.15 (version 5 per FIRST research page), per the CSV header on 2026-11-10
Updated
Daily, typically a minute or two after 13:30 UTC
Answers
The likelihood of exploitation in the wild in the next 30 days, across all organizations

What it is

FIRST describes EPSS as a data-driven machine-learning model that estimates the probability that a published CVE will be exploited in the wild in the next 30 days. It publishes a probability between 0 and 1, with ranking percentiles, every day for every CVE, free through CSV, an API and a GitHub repository.

The CSV file carries its own version on the first line. On 2026-11-10 the header read #model_version:v2026.06.15, with a score_date of 2026-11-10. The FAQ states that retraining has occurred approximately annually and is triggered by sustained drift toward the random baseline, measured by comparing scores from 30 days earlier with the exploitation seen in the following 30 days.

The data sit in a public repository as well, which allows historical scores to be pulled by date. This site's ledger uses that: the FIRST API call returns the score and percentile for a given CVE and date, and the ledger asks for the dateAdded of each KEV entry. The result is the score as published that day, not the later score.

Inputs and method

The How It Works page states that each daily record is built from approximately 2,850 features, including exploit code and offensive tooling, and that EPSS draws on nearly a dozen separate sources of exploitation activity across four sensor types. FIRST names VulnCheck KEV, the Shadow Server Foundation and LevelBlue as data partners, among others.

Two details matter to an auditor. Exploitation activity trains the model but does not drive daily scoring, so a confirmed event does not move tomorrow's score directly. And the training data and operational pipeline are not published, because data partners share telemetry under agreements that prohibit redistribution. The inputs, architecture and held-out performance are documented; the pipeline is not.

Reading probability and percentile

The FIRST article on probability and percentile bins gives the working example. A score of 0.10 sits in the 88th percentile, the top 12 percent of 170,000 or more scored vulnerabilities, even though 10 percent is a low absolute probability. Out of vulnerabilities at 10 percent, about 1 in 10 is expected to be exploited in the next 30 days.

FIRST's guidance is to communicate the probability as the EPSS score and to show the percentile beside it, for example 15.3% (92nd). The EPSS SIG does not bin scores into labels, because bin cutoffs are subjective and compress scores that differ by 0.002 into opposite categories. A program that adds its own bins should record the cutoff and who set it.

A related point from the same article: percentiles are calculated over every published CVE, so the percentile for a subset of vulnerabilities in your environment changes while the probability does not. If a report shows percentiles, state the population they were computed over.

What the KEV Lag Ledger shows

Among the 318 KEV additions in the ledger window, 260 had an EPSS value from the FIRST API on the day CISA listed them. The other 58 had none that day, which is typical of a CVE only a day or two old. For the 260, the median EPSS on listing day was 0.0066 (0.006555), and 75.8 percent (197) were below 0.1.

The upper end is large too: 63 listings (24.2 percent) were at or above 0.1, 34 (13.1 percent) at or above 0.5, and 75 (28.8 percent) in the 90th percentile or higher. A threshold of 0.1 on listing day would have flagged one in four of these confirmed vulnerabilities and missed three of every four. This is a listing-day snapshot and does not measure the 30 days before it.

Documented limits

The How It Works page lists its own limits. Scores are population-level and not tailored to an organization or technology stack. Newly disclosed vulnerabilities have sparse signals, so a score on day one is thin. Training labels reflect only exploitation that data partners can observe. FIRST also states that a low score is not a declaration of safety: across 10,000 vulnerabilities at 5 percent, about 500 exploitation events are expected.

The Using EPSS page adds two rules for a policy. Do not multiply EPSS by an ordinal such as CVSS to produce a combined score; FIRST calls that score laundering. And when a vulnerability appears on KEV, treat it as actively exploited regardless of its EPSS score.

The FAQ also answers a question auditors ask: how often scores are generated. Daily, typically within a minute or two after 13:30 UTC. A dashboard that refreshes EPSS weekly is showing stale values, and a policy should state the refresh interval it actually uses.

Using it in a ranking

Use the probability for the large population KEV does not cover. The Using EPSS page puts the base rate of exploitation at roughly 2 to 3 percent, and note that half of published vulnerabilities score above about 0.7 percent, so a threshold needs to be set from the distribution, not from intuition.

Pair it with KEV: KEV answers what has already happened, EPSS what is likely next. See the pages comparing EPSS with CVSS and with KEV for the combinations, and the SSVC page for how CERT/CC and CISA propose feeding EPSS into a decision tree.

Thresholds are a program decision with a cost on each side. The How It Works page frames the choice as coverage against efficiency: a lower threshold flags more vulnerabilities and captures more of the exploited ones, at the cost of effort spent on vulnerabilities never exploited. Record the threshold, the date it was chosen and the coverage you accepted.

Comparisons with this system

Sources

  1. FIRST, EPSS home, read 10 Nov 2026
  2. FIRST, How EPSS works (features, limitations), read 10 Nov 2026
  3. FIRST, Using EPSS (thresholds, misuses, KEV guidance), read 10 Nov 2026
  4. FIRST, Probability, percentile and bins, read 10 Nov 2026
  5. FIRST, EPSS FAQ (daily publication, retraining), read 10 Nov 2026
  6. EPSS daily CSV (model version header, 2026-11-10), read 10 Nov 2026

Last reviewed 10 Nov 2026.