Datasets:

drumshadow
/

vulnscan-pr-security

license: cc-by-4.0
task_categories:
  - text-classification
language:
  - en
tags:
  - security
  - vulnerability-detection
  - code-review
  - benchmark
pretty_name: vulnscan PR security review benchmark

vulnscan: PR security review benchmark (v1: Python + JS/TS)

A benchmark for AI security review of code changes. Every advisory was published on or after 2026-07-01. An advisory date does not bound when the fix became public: in the test split, 66 of 163 fixes were committed before that date. Use subsets/test_nc_fix_after_cutoff_pairs.json (also evidence/ in the GitHub repo) for the 97 pairs whose fix came after it. No claim is made about any model's training data. The split unit is the repository, and a build-time leak check aborts on any repository, commit or advisory overlap between splits (0 shared resolved SHAs across all splits).

This release contains metadata only: advisory ids, repository names, commit SHAs, file lists, view windows, truth lines and boolean leak flags, enforced by a field-level allow-list at staging. No third-party source code is redistributed. rebuild_from_release.py (in the repo) regenerates the reversed-fix diffs from the SHAs (all 1,259 byte-identical on our copy). New-code items are pointers read from git.

Two framings

newcode/ (headline). Each vulnerability appears as a pair: the vulnerable version and the patched version of the same files, both presented as added code. Ordinary commits from the same repositories are included as negatives. An 8-feature surface-statistics probe scores AUC 0.51 (test) / 0.54 (dev) on the rendered prompts, so the deletion cue is gone. That does not rule out every other surface signal. Metric: pair accuracy, meaning find the vulnerability in the vulnerable version and don't flag the fix site in the patched one.
regression/ (secondary, shortcut warning). The reverse of each fix commit, presented as a PR. These diffs are mostly deletions, and the same probe scores AUC 0.95 on them (vs ordinary commits and fixes; 0.94 vs ordinary commits alone). On the same advisories, frontier models' line-window recall is 3–4× higher here than on the new-code framing (the framings also differ in context and instructions). It is included to make that measurable.

Split	Pairs (new-code)	Ordinary commits	Repos
dev	164	298	142
test	163	305	134

Fields (per item)

truth: lines (in head) the fix changed.
fix_site: lines (in the patched version) the fix added.
leak: residual answer-leak flags (deleted code naming the bug). Report results with and without flagged items.

Scoring

Line window: right file, within ±5 lines of a truth line, right CWE family (vs/cwe.py).
Semantic judge: a fixed rubric decides whether a finding is the advisory's vulnerability (same root cause and component, any line). A non-specialist hand check (the author, blind to reviewer and verdict) agreed with the judge on 20 of 28 decided rows (Cohen's κ 0.41); in 7 of the 8 disagreements the judge was the stricter one (evidence/judge-hand-check.md). That is not an expert validation.

Reference results used one sample: all 163 test pairs plus 80 ordinary commits (406 items). Semantic pair accuracy on the 97 fix-after-cutoff pairs: Codex gpt-6-astra 33% (24–43), Claude Opus 5.5 24% (16–33), local Qwen3.6-35B-A3B harness 13% (7–21), semgrep 5% (1–10). On all 163 pairs: 40%, 32%, 20%, 7%.

Known limitations

Ordinary commits are not audited. A few may contain real vulnerabilities and count as false alarms.
GHSA CWE labels are coarse, so scoring uses CWE families.
The new-code framing presents whole files (or ±150-line windows) as added code. It is a controlled proxy for PR review, not a replay of real PRs.

Sources and license

Advisory metadata from the GitHub Advisory Database (CC-BY-4.0) and NVD. Repositories are referenced by commit SHA under their own licenses; no code is included. This metadata: CC-BY-4.0.