README.md

---
license: cc-by-4.0
language:
  - en
pretty_name: Red Hat VEX Fixed CVEs with Upstream Fix Commits
tags:
  - security
  - vulnerabilities
  - cve
  - vex
  - code-security
  - benchmark
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data.parquet
---

# Red Hat VEX Fixed CVEs with Upstream Fix Commits

One row per CVE that Red Hat has **fixed** in at least one shipped product
(per Red Hat's CSAF VEX data), joined against CVE.org's `cvelistV5` records
to resolve the **upstream git commit(s)** that fixed the issue. Only CVEs
with at least one resolved upstream fix commit are included.

Intended use: benchmarking the ability of models and tools to find real
security issues in code. For each row with a resolved fix commit, the parent
commit (`<upstream_fix_commit>^`) of `upstream_repo_url` is a known-vulnerable
code state, and the commit diff is ground truth for the flaw and its fix.

## Fields

| Field | Description |
|---|---|
| `cve` | CVE identifier |
| `title` | Red Hat flaw title (`component: short description`) |
| `component` | Component name parsed from the title |
| `severity` | Red Hat aggregate severity (Low / Moderate / Important / Critical) |
| `cwe_id`, `cwe_name` | CWE assigned by Red Hat |
| `cvss_v3_score`, `cvss_v3_vector` | Highest CVSS v3 base score across affected products |
| `description` | Red Hat vulnerability description |
| `summary` | Red Hat topic/summary line |
| `statement` | Red Hat statement (context / applicability), when present |
| `public_date` | Date the issue became public (ISO 8601) |
| `discovery_date` | Red Hat discovery/report date |
| `bugzilla_id` | Red Hat Bugzilla identifier |
| `cisa_kev` | True if referenced in CISA Known Exploited Vulnerabilities catalog |
| `rh_advisories` | Red Hat advisories (RHSA/RHBA/RHEA) that shipped a fix |
| `rh_fixed_packages` | Fixed package NEVRAs (capped at 50 per row) |
| `references` | Reference URLs from the VEX record |
| `upstream_fix_commit` | Primary upstream fix commit SHA (first resolved) |
| `upstream_repo_url` | Upstream repository URL for that commit |
| `upstream_fix_commits` | All resolved commits: `{repo_url, commit, ref_url, source}` |
| `upstream_affected_versions` | Affected upstream version ranges (cvelistV5 CNA data, or `<= <tag>` derived from git when the CNA had none) |
| `upstream_fixed_versions` | Fixed upstream versions (cvelistV5 CNA data, or the git-derived tag when the CNA had none) |
| `upstream_fixed_version_git` | First upstream release tag whose history contains the fix commit (derived by cloning the repo and running `git tag --contains`) |
| `upstream_last_affected_version_git` | Newest release tag in the fix commit's ancestry — the latest release that definitely does **not** contain the fix |
| `version_derivation_status` | `ok`, `fix_unreleased` (no tag contains the commit yet), `commit_not_found`, `clone_failed`, `no_clone_url` (non-git upstream, e.g. Mercurial) |
| `resolution_source` | How the commit was resolved: `redhat_vex` (commit URL present in Red Hat's own references) or `cvelist` (CVE.org record) |

## Construction

1. Downloaded the full Red Hat CSAF VEX archive from
   `https://security.access.redhat.com/data/csaf/v2/vex/` and kept only CVEs
   with a non-empty `product_status.fixed`.
2. Extracted commit URLs (GitHub, GitLab, cgit, gitweb styles) from the VEX
   references.
3. Joined against the CVE.org `cvelistV5` bulk dataset for additional commit
   references (including kernel.org CNA `versionType: git` ranges) and
   upstream affected/fixed versions.

4. Dropped rows without a resolvable upstream commit.
5. Cloned each upstream repository (972 unique repos, treeless bare clones)
   and derived release-tag boundaries for every fix commit:
   `upstream_fixed_version_git` = version-lowest stable tag containing the
   commit; `upstream_last_affected_version_git` = version-highest stable tag
   in the commit's ancestry. Verified against known ground truths
   (Heartbleed → `OpenSSL_1_0_1g`, SACK Panic → `v5.2`, etc.) and
   cross-checked against CNA-declared fixed versions (96% numeric agreement
   on the comparable subset).

## Statistics

- 110,635 Red Hat VEX records scanned; 25,428 CVEs with fixed status, of
  which **7,664** have at least one resolved upstream fix commit and form
  this dataset (1,025 resolved from Red Hat's own references, 6,639 from
  CVE.org)
- Commit-resolution rate by era: ~15–25% of fixed CVEs pre-2021, 50–60% for
  2022–2025 (driven largely by the Linux kernel CNA publishing fix commits) —
  so coverage skews toward recent CVEs
- Top components: Linux kernel (~4,200), OpenSSL, PHP, tcpdump, ghostscript,
  FreeRDP
- Version coverage: 7,011 rows (91.5%) with git-derived fixed version, 7,076
  with git-derived last-affected version; `upstream_fixed_versions` populated
  for 7,057 rows after combining CNA and git sources

## Caveats

- The git-derived version fields describe the **exact fix commit's** history:
  for projects that cherry-pick fixes onto release branches (new SHA), the
  first tag containing the mainline commit can be later than the first
  *release* that shipped a backported fix. E.g. a tcpdump fix cherry-picked
  into 4.9.2 but committed on master shows `upstream_fixed_version_git:
  tcpdump-4.99.0`. The CNA-provided `upstream_fixed_versions` field retains
  the release-branch view where available.
- Tag names are reported verbatim (`v6.8`, `OpenSSL_1_0_1g`,
  `curl-7_88_1`, ...) — normalize per-project if you need semver.
- Commit SHAs are taken from published references; a sampled subset was
  verified to exist in the claimed repositories, and every row with
  `version_derivation_status: ok` had its commit positively resolved in a
  clone of the upstream repo. Multi-commit fixes may still be incomplete.
- `upstream_repo_url` is derived from the commit URL and may point at a web
  UI path (cgit/gitweb) rather than a clean clone URL.
- Affected/fixed version data comes from CVE.org CNA containers and varies in
  quality per CNA.
- A "fixed" status here means Red Hat shipped a fix in at least one product;
  other products may remain affected or out of support scope.
- For model benchmarking, filter on `public_date` relative to a model's
  knowledge cutoff to control for memorization.

## Sources and license

- Red Hat CSAF VEX data © Red Hat, Inc., licensed
  [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).
- CVE records from [CVE.org](https://www.cve.org) (CVE Program), used per the
  [CVE Terms of Use](https://www.cve.org/Legal/TermsOfUse).

This dataset is a transformation of those sources and is distributed under
CC BY 4.0. Snapshot dates: VEX archive 2026-07-27, cvelistV5 2026-07-30.