EPSS and exploit likelihood

The Exploit Prediction Scoring System: what the probability and percentile actually mean, where the data comes from, how to pull scores from the FIRST API, how to combine EPSS with CVSS without double counting, and the limits that stop it being a priority on its own.

FREEBy ·Lesson·9 min read·Intermediate → Advanced
EPSSexploit predictionFIRSTrisk-based vulnerability managementpatch prioritisation

01What EPSS measures

✓What you will learn
  • Read an EPSS probability and percentile correctly
  • Pull scores from the FIRST API and join them to scanner output
  • Combine EPSS with CVSS as two different axes, likelihood and impact
  • Know the limits: what EPSS cannot see and when to override it

EPSS (Exploit Prediction Scoring System), run by a FIRST special interest group, estimates the probability that exploitation activity against a CVE will be observed in the next 30 days. Scores are recalculated daily for every published CVE, from a model trained on observed exploitation attempts reported by data partners (intrusion detection, honeypots and similar sensors) and features of the vulnerability: vendor, CWE, references, whether exploit code is public, mentions in security lists, and age.

FieldExampleMeaning
epss0.94310About a 94% chance of observed exploitation activity in the next 30 days
percentile0.99930This CVE scores higher than 99.93% of all scored CVEs
date2026-11-16The model run that produced the score; scores move daily

The probability is the number to use for decisions; the percentile is a ranking. Most CVEs have very low probabilities, so the distribution is extremely skewed: a CVE with a probability of a few per cent can sit in the top tenth by percentile. Reporting only the percentile makes ordinary vulnerabilities sound alarming.

iExploitation activity, not compromise

EPSS predicts that someone, somewhere, will attempt exploitation that partners' sensors observe. It does not predict that you will be compromised, and it says nothing about impact. That is why it sits beside CVSS, not instead of it.

02Getting scores

The FIRST EPSS API and daily CSVbash
# One CVE
curl -s "https://api.first.org/data/v1/epss?cve=CVE-2021-44228"

# Several at once, and the history for one
curl -s "https://api.first.org/data/v1/epss?cve=CVE-2024-3400,CVE-2023-4966,CVE-2019-0708"
curl -s "https://api.first.org/data/v1/epss?cve=CVE-2024-3400&scope=time-series"

# Highest-probability CVEs today
curl -s "https://api.first.org/data/v1/epss?epss-gt=0.5&order=!epss&limit=20"

# Full daily file for bulk joins: a gzip CSV of every scored CVE, linked from first.org/epss/data
curl -sLO "$EPSS_DAILY_CSV_URL"     # set from the link on the EPSS data page
Python: enrich a scanner export with today's EPSSpython
import csv, gzip, io, urllib.request

def load_epss(path: str) -> dict[str, float]:
    with gzip.open(path, "rt") as fh:
        rows = (line for line in fh if not line.startswith("#"))   # first line is a model comment
        return {r["cve"]: float(r["epss"]) for r in csv.DictReader(rows)}

epss = load_epss("epss_scores-current.csv.gz")

with open("scanner_export.csv", newline="") as fh, open("findings_epss.csv", "w", newline="") as out:
    rd = csv.DictReader(fh)
    wr = csv.DictWriter(out, fieldnames=[*rd.fieldnames, "epss"])
    wr.writeheader()
    for row in rd:
        row["epss"] = epss.get(row["cve"], "")          # blank means not scored, not zero
        wr.writerow(row)
!Missing is not zero

Very new CVEs and rejected ones may have no score. Write blank, not 0.0, so a fresh critical does not sort to the bottom of the queue the day it is published, which is exactly when attackers move.

03Combining EPSS with CVSS

CVSS answers how bad if exploited; EPSS answers how likely to be exploited soon. Plot them as two axes rather than multiplying them into one opaque number, and agree thresholds with the owners in advance.

EPSS high (for example 0.1 and above)EPSS low
CVSS High / CriticalAct now: likely and damagingSchedule: damaging if it happens; patch in the normal cycle, watch for EPSS or KEV changes
CVSS Medium / LowInvestigate: often a chaining step or a widely deployed product; check exposureBacklog: fix with routine updates

FIRST's own analysis shows the point of this. Patching everything with CVSS 7 or above means working through a large share of all CVEs to cover most exploited ones; an EPSS threshold reaches similar coverage with far fewer patches, at the cost of missing some. No threshold is free: you trade coverage (share of exploited CVEs you fixed) against efficiency (share of your fixes that were for exploited CVEs). Pick the threshold knowing which you are giving up.

▭On the whiteboard · Choosing an EPSS threshold5 points
01
capacity
How many fixes can the team ship per cycle?
02
coverage
What share of exploited CVEs must the queue catch?
03
exposure
Lower the threshold for internet-facing assets
04
override
KEV listing or vendor 'exploited in the wild' trumps any threshold
05
review
Re-run daily: probabilities jump when exploit code is published

04Limits and failure modes

  • Visibility bias: EPSS learns from what partners' sensors see. Exploitation that sensors rarely observe (targeted attacks on niche appliances, flaws needing a user to open a file, internal-only protocols) is under-predicted.
  • No asset context: the score is the same for an internet-facing VPN and a lab box. Exposure and criticality are yours to add.
  • Lag: scores jump after public exploit code or the first observed attempts. A day-zero critical on an edge device can sit at a low probability for its first day or two.
  • Model versions change: EPSS v4 was released in March 2025, and each new model shifts scores. Record the model version and date with any decision you might have to defend.
Test yourself

A CVE has EPSS probability 0.03 and percentile 0.91. A manager reads it as "top 10%, must patch now". What is the accurate reading?

05Putting EPSS to work

EPSS earns its place when it runs every day without anyone thinking about it. Wire it into the enrichment that builds the queue, and use it mostly to decide what can safely wait.

  1. Pull the daily scores and join them to open findings by CVE, storing the probability, percentile and model date.
  2. Flag findings whose probability crossed your threshold since yesterday; those are the changes worth a message.
  3. Combine with exposure: apply a lower threshold to internet-facing and critical assets.
  4. Let KEV listings and vendor exploitation notices override any EPSS value.
  5. Review the threshold quarterly against what was actually exploited and what the team could ship.
“How to say it in the room

"EPSS is the probability that exploitation activity is seen in the next 30 days. I use the probability, not the percentile, as a likelihood axis next to CVSS impact, I override it with KEV and vendor exploitation notices, and I lower the bar for anything internet facing because EPSS knows nothing about my exposure."

◎Interview radar

A popular question is "CVSS or EPSS?". The strong answer is both, as different axes, plus KEV for confirmed exploitation and asset context for exposure. Being precise about probability versus percentile and the 30-day window separates people who have used EPSS from people who have read about it.

Q1Can EPSS replace threat intelligence?+

No. It is a population-level prediction. Intelligence that an actor targeting your sector is exploiting a specific CVE should override it immediately.

Q2Why might a high-CVSS CVE have a tiny EPSS?+

It may need local access, a rare configuration or user interaction, have no public exploit, or affect a product few attackers bother with.

Q3How often should you refresh scores?+

Daily, ideally as part of the enrichment that builds the queue, with alerts when an open finding crosses your threshold.

Finished this one?
Tick it and it shows as done on the roadmap and the course.