Vuln Prioritisation

Predicting which new CVEs get exploited

A public, running test of Vulnerability Prioritisation v1 (VPv1). Every morning it scores the CVEs published in the previous 24 hours and names the ones it expects to be exploited within 90 days. Each day’s list is published here as it stands and never edited afterwards, so the predictions can be checked against what actually happens.

Days running
—
…
CVEs scored
—
…
VPv1 top 5%: CVEs needed
—
…
EPSS v5: CVEs needed
—
…
Exploited so far
—
…

What VPv1 is

More than 72,000 CVEs have been published so far in 2026, and well under 1% of them are ever exploited. Nobody can patch everything at once, so the useful question on the day a CVE comes out is which ones to deal with first.

VPv1 answers that question. It looks at each new CVE using only information that is public on the day it is scored and ranks every CVE. It flags the top 10% of that ranking, the CVEs it expects are most likely to be exploited within 90 days, and marks each one by how high it ranks: top 2%, top 5% or top 10%. Which band to work to is your choice: a smaller band means fewer CVEs to review but fewer exploits caught, so pick the one that fits your team’s capacity and appetite for risk.

CVEs that are already known to be exploited when they are published, such as a vendor disclosing a zero-day that is under active attack, are left out. There is nothing to predict for those, and counting them would make the results look better than they are.

How well it performed in testing

VPv1 was tested on CVEs published between 17 March 2025 and 31 May 2026, the period in which EPSS v4 was the current version of EPSS. Each CVE was scored using only what was public on the day after it was published, and the model never saw outcomes from later than the point it was scoring. Zero-days already exploited at disclosure are excluded from both sides.

The table shows how many CVEs a team would need to review each year to catch a given share of the CVEs exploited within 90 days.

Share of exploited CVEs caught50%70%80%90%
EPSS v43,55413,58319,58635,059
VPv15691,7022,8746,712
Difference2,985 fewer (6×)11,881 fewer (8×)16,712 fewer (7×)28,347 fewer (5×)

CVEs reviewed per year, annualised from 17 March 2025 to 31 May 2026 (about 56,500 CVEs and 93 exploited within 90 days a year).

VPv1’s top 10% is split into three bands, so a team can work to the band that suits it or work down them in order:

BandCVEs per yearPer dayExploited CVEs caughtEPSS v4, same number of CVEsEPSS v4 CVEs needed for the same catch
Top 2%1,1303.163%31%7,509
Top 5%2,8257.779%45%19,292
Top 10%5,65015.587%55%30,233

Each band includes the ones above it. Backtest, 17 March 2025 to 31 May 2026, zero-days already exploited at disclosure excluded.

Those are backtest results, worked out after the fact on historical data. The live test below is the real check.

How the live test works

  • Every morning between 8am and 9am AWST, VPv1 scores every CVE published to the CVE List in the previous 24 hours.
  • Every CVE scored since the start of the test is scored again with whatever has become public since. A CVE can be flagged days after it was published if new information appears, such as public exploit code, exploitation of another flaw in the same product, or (from 11 October) a likely chain partner already flagged.
  • A CVE is flagged when its score ranks in VPv1’s top 10%, and is marked top 2%, top 5% or top 10%. The cut-offs are set each day from data available that morning, not from later results. If new information moves a CVE into a higher band, the move is recorded on the day it happens.
  • Each day’s results are published as a separate dated file and are never changed afterwards. Each file records a SHA-256 fingerprint of the previous day’s file, so editing any earlier day would break the chain.
  • The page and each daily file are captured by the Internet Archive’s Wayback Machine, which gives an independent record of what was predicted and when.

Changes to how exploitation is counted

Changes to how exploitation is counted were initially made in October, after the experiment had started. Those changes have been revoked so that the results are not perceived as being altered to favour VPv1.

How it is scored

A CVE counts as exploited on the earliest of two dates: the first credible public report of exploitation in the wild, or its addition to the CISA KEV catalog. Credible reports are vendor advisories, government advisories and named security research or incident reporting, the same standard used for the rest of this site. Reported exploitation attempts from a named source count. A priority or severity rating, or a statement that a flaw is likely to be exploited, does not.

A prediction only counts if VPv1 had flagged the CVE before its exploitation became public (for a band, if the CVE had reached that band or a higher one by then). Being flagged after the news broke earns nothing. When a CVE is flagged on the same day its exploitation becomes public, the time of day decides. The CVE was flagged at the time of that morning’s run, which is shown in each daily file. A report or KEV listing whose time is not known counts from the start of its day (00:00 UTC), so an unknown time never favours VPv1.

Reports of exploitation are searched for every day across every CVE in the test, not just the flagged ones, so neither VPv1 nor EPSS is checked more closely than the other. Each exploited CVE below links to the report its date comes from.

Each flagged CVE carries one of four statuses:

  • Open: published less than 90 days ago, with no public exploitation yet.
  • Exploited within 90 days: exploitation became public within 90 days of publication.
  • Not exploited within 90 days: 90 days passed with no public exploitation.
  • Exploited after 90 days: exploitation became public later than 90 days after publication.

A CVE later found to have been exploited on or before the day it was published, or before the morning run that first scored it, is marked Excluded and left out of every result, because there was nothing to predict.

The results are reported in three ways:

  • Coverage: the share of all CVEs exploited within 90 days that VPv1 had flagged (in its top 10%, or in the band being measured) before exploitation became public.
  • CVEs reviewed per exploit caught: how many CVEs a team working through VPv1’s top 10% (or one band of it) reviewed for each exploited CVE it caught in time.
  • Comparison with EPSS: each morning, every CVE in the test that is under 90 days old and not yet known to be exploited is ranked by that day’s EPSS score. A CVE joins EPSS’s top 10%, 20%, 30% or 50% on the first morning it ranks there and stays in, just as CVEs stay flagged by VPv1. CVEs with the same EPSS score are counted together, since EPSS gives no way to choose between them. Each EPSS list is scored exactly like VPv1’s. The results also show the smallest top share of EPSS’s ranking that would have caught every exploit VPv1 caught in time, and how many CVEs that list held. VPv1’s list is measured the same way: its CVEs are ranked by band and, within a band, by VPv1’s score, and the list is cut at the last exploit it caught in time. EPSS released version 5 in 2026, so the live comparison is against EPSS v5, the version published each day. The model version is recorded in every daily file.

Results for a CVE become final 90 days after it is published, so the first complete numbers arrive about three months after the test starts. Until then the figures above count what has happened so far.

Results so far

VPv1’s top 2%, top 5% and top 10% (everything VPv1 flags) against the top 10%, 20%, 30% and 50% of EPSS v5’s ranking of the same CVEs, all scored the same way. VPv1’s top 5% includes its top 2%, and a CVE counts in a band from the morning it first reached that band or a higher one. The question is how far down EPSS’s ranking a team would have to go to catch what VPv1 catches. “CVEs in both” counts CVEs on both lists: for each VPv1 row, those also in EPSS’s top 10%; for each EPSS row, those also in VPv1’s top 10%. A list can hold more than its share because a CVE stays on once it has ranked there. The numbers keep changing until each CVE’s 90 days are up.

CVEs flaggedCVEs in both% CVEs in bothExploits caught in timeCoverageCVEs reviewed per exploit caught
Loading…

Exploited CVEs so far

Every CVE in the test whose exploitation in the wild has become public, whether or not it was flagged. A flag only counts if it came before the first public report or KEV listing. For EPSS, the column shows the smallest of its top 10/20/30/50% that had the CVE in time, and how far down its ranking the CVE was at best before exploitation became public. Times are UTC.

CVEVendorProductPublishedExploitation publicCISA KEVVPv1EPSS v5
Loading…

VPv1’s flagged CVEs

Grouped by the day each CVE was flagged, newest first.

Loading…

Daily files

Every run, as published. Each file contains the SHA-256 of the one before it.

Run dateCVEs publishedFlaggedFileSHA-256
Loading…