A static benchmark goes stale the moment the tools it measures improve. Every release is identified by a deterministic manifest hash, so old results stay valid and versions compare side by side.
A manifest hash is computed from three inputs: the content hash of the ground-truth directory, the pinned commit SHA of each repository, and the hash of the default prompt used for LLM scanners. Any change to labels, code, or prompt produces a new version; scores are always reported against a named manifest, and the dashboard surfaces historical comparisons.
Type 2 targets introduce realistic, low-density vulnerability profiles — a higher bar for precision than intentionally vulnerable apps.
Hypothesis: LLM-based scanners may detect vulnerabilities in LLM-generated code disproportionately better than in human-authored code. RealVuln tags every repository's authorship — the method runs each scanner against matched human-authored and LLM-generated pairs and compares the deltas. The result would be informative independent of which scanner ranks highest.
RealVuln is maintained as a community resource. Transparency is treated not as a defense against bias but as an invitation to find it. Contribution paths are documented in the repository README.
Place results in Semgrep-compatible JSON and score. Unknown slugs fall back to the default parser; other formats need a parser class registered in parsers/.
Author a ground-truth manifest following the schema, validate it, and pin a commit SHA. At least one false-positive trap is required per five real findings.
Every label ships with evidence and rationale. Open an issue to dispute one; corrections produce a new manifest version without invalidating prior results.
# validate ground-truth schemas python validate_gt.py # score one repo against every scanner python score.py --repo realvuln-pygoat --all-scanners # regenerate the interactive dashboard python dashboard.py --scanner-group all
If RealVuln is useful in your work, please cite the paper. All artifacts are released under the MIT license.
@misc{pellew2026realvuln, title = {RealVuln: Benchmarking Rule-Based, General-Purpose LLM, and Security-Specialized Scanners on Real-World Code}, author = {Pellew, John and Raza, Faizan}, year = {2026}, eprint = {2604.13764}, archivePrefix = {arXiv}, primaryClass = {cs.CR}, url = {https://arxiv.org/abs/2604.13764} }