KEV additions without a day-before EPSS score, and the day-before EPSS percentile of scored additions, in the EPSS v3 and v4 eras

Of CISA Known Exploited Vulnerabilities (KEV) additions, 28.3% (117 of 414) in the EPSS v3 era (added 2023-03-10 to 2025-03-13) and 23.8% (76 of 319) in the v4 era (2025-03-18 to 2026-06-15) had no EPSS score in the file dated the day before listing. v4 minus v3, -4.4 percentage points (95% CI -13.8 to +4.8). Of these 193 CVEs, 188 carry the year of addition in their id, and post hoc all 193 were in a later downloaded file, so the share reflects how new the CVEs were more than the model. The v3 share exceeds the 25% limit of stop rule 1 of a privately pre-specified plan, so everything else is descriptive only. Among scored additions, 37.4% (v4 era) against 26.3% (v3 era) were at or above the 90th percentile of their day-before file (+11.2 points, 95% CI +2.4 to +19.9), and the median percentile was 0.721 against 0.427 (+0.294, +0.056 to +0.441); intervals resample addition days within each era. Post hoc, with the unscored additions of one era at percentile 0 and of the other at 1, the all-additions median difference runs from -0.433 to +0.690, so its sign is not fixed. Post hoc, of 100 v4-era additions scored in both the last v3 file (2025-03-16) and the first v4 file (2025-03-17), 36.0% and 62.0% were at or above 0.90; 66.0% rose, and the median per-CVE change was +0.032; v4 may have been trained on these CVEs' own exploitation. Post hoc, in v4-era files the top decile also held more recent CVEs outside KEV: of the non-KEV CVEs carrying the addition year in their id, a median of 0.80% were in it, against 0.31% in v3-era files. Post hoc, the median ranged from 0.291 to 0.593 between half-years inside the v3 era alone. The eras differ in which CVEs CISA added, in the scored set behind the percentile and in how the files rank tied scores, so the difference coincides with the model switch and is not attributed to it. Data: KEV catalogVersion 2026.11.18 (CC0); EPSS daily files (EPSS by FIRST). The plan was committed to a private repository before the download; a reader cannot verify its time. The record holds the note (PDF and Markdown) and a supplement archive: the frozen plan and the feasibility-check ("gate") files (internal references replaced), the deviations log, the pre-review result record (edited as its manifest row lists), the download manifests with the URL, size and sha256 of the 424 day-before EPSS files, the KEV snapshot record, the per-addition presence and file header tables, the stop-rule and detector-check files, the registered estimates, the download, stop-rule and outcome code, the checks added for the note (paper_checks.py, paper_checks.json) and after the reviews by AI sessions (post_review_checks.py, post_review_checks.json, post_review_fetch.py, post_review_fetch.json), a manifest with the sha256 of every file and every change made for the deposit, and a trace of every number in the note. The EPSS daily files and the KEV feed are not redistributed, and the record holds no EPSS score or percentile tied to a CVE id and no plot of them: the code rebuilds the per-addition percentile table and its distribution plot from the public files. EPSS scores: EPSS by FIRST (Forum of Incident Response and Security Teams, https://www.first.org/epss), generated and published by Empirical Security. KEV: CISA Known Exploited Vulnerabilities catalog, CC0 1.0. Plan, code, analysis and text by Claude (an AI model by Anthropic) for the author; no human has checked the figures; the human author is responsible for the deposit.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24401267
Primary Topic
Information and Cyber Security
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

KEV additions without a day-before EPSS score, and the day-before EPSS percentile of scored additions, in the EPSS v3 and v4 eras

Keisuke Watanabe
Zenodo (CERN European Organization for Nuclear Research)
Information and Cyber Security
article

KEV additions without a day-before EPSS score, and the day-before EPSS percentile of scored additions, in the EPSS v3 and v4 eras

Keisuke Watanabe
article en

Abstract

Of CISA Known Exploited Vulnerabilities (KEV) additions, 28.3% (117 of 414) in the EPSS v3 era (added 2023-03-10 to 2025-03-13) and 23.8% (76 of 319) in the v4 era (2025-03-18 to 2026-06-15) had no EPSS score in the file dated the day before listing. v4 minus v3, -4.4 percentage points (95% CI -13.8 to +4.8). Of these 193 CVEs, 188 carry the year of addition in their id, and post hoc all 193 were in a later downloaded file, so the share reflects how new the CVEs were more than the model. The v3 share exceeds the 25% limit of stop rule 1 of a privately pre-specified plan, so everything else is descriptive only. Among scored additions, 37.4% (v4 era) against 26.3% (v3 era) were at or above the 90th percentile of their day-before file (+11.2 points, 95% CI +2.4 to +19.9), and the median percentile was 0.721 against 0.427 (+0.294, +0.056 to +0.441); intervals resample addition days within each era. Post hoc, with the unscored additions of one era at percentile 0 and of the other at 1, the all-additions median difference runs from -0.433 to +0.690, so its sign is not fixed. Post hoc, of 100 v4-era additions scored in both the last v3 file (2025-03-16) and the first v4 file (2025-03-17), 36.0% and 62.0% were at or above 0.90; 66.0% rose, and the median per-CVE change was +0.032; v4 may have been trained on these CVEs' own exploitation. Post hoc, in v4-era files the top decile also held more recent CVEs outside KEV: of the non-KEV CVEs carrying the addition year in their id, a median of 0.80% were in it, against 0.31% in v3-era files. Post hoc, the median ranged from 0.291 to 0.593 between half-years inside the v3 era alone. The eras differ in which CVEs CISA added, in the scored set behind the percentile and in how the files rank tied scores, so the difference coincides with the model switch and is not attributed to it. Data: KEV catalogVersion 2026.11.18 (CC0); EPSS daily files (EPSS by FIRST). The plan was committed to a private repository before the download; a reader cannot verify its time. The record holds the note (PDF and Markdown) and a supplement archive: the frozen plan and the feasibility-check ("gate") files (internal references replaced), the deviations log, the pre-review result record (edited as its manifest row lists), the download manifests with the URL, size and sha256 of the 424 day-before EPSS files, the KEV snapshot record, the per-addition presence and file header tables, the stop-rule and detector-check files, the registered estimates, the download, stop-rule and outcome code, the checks added for the note (paper_checks.py, paper_checks.json) and after the reviews by AI sessions (post_review_checks.py, post_review_checks.json, post_review_fetch.py, post_review_fetch.json), a manifest with the sha256 of every file and every change made for the deposit, and a trace of every number in the note. The EPSS daily files and the KEV feed are not redistributed, and the record holds no EPSS score or percentile tied to a CVE id and no plot of them: the code rebuilds the per-addition percentile table and its distribution plot from the public files. EPSS scores: EPSS by FIRST (Forum of Incident Response and Security Teams, https://www.first.org/epss), generated and published by Empirical Security. KEV: CISA Known Exploited Vulnerabilities catalog, CC0 1.0. Plan, code, analysis and text by Claude (an AI model by Anthropic) for the author; no human has checked the figures; the human author is responsible for the deposit.

Zenodo (CERN European Organization for Nuclear Research)
Information and Cyber Security
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon