| # Responsible Disclosure & Vulnerability Handling |
|
|
| AppSecBench is a benchmark of *known, catalogued* weakness classes. It does not contain |
| zero-days or undisclosed vulnerabilities. The guidance below applies to the project itself and to |
| researchers who build on the dataset. |
|
|
| ## Reporting a problem with the dataset |
|
|
| If you find that a record is incorrect, mislabeled, or inadvertently reproduces copyrighted or |
| sensitive material, please disclose privately: |
|
|
| 1. Open a private security advisory on the Hugging Face dataset repository, **or** |
| 2. Email the maintainer (see `SECURITY.md`) with the `benchmark_id`(s) and the issue. |
|
|
| We aim to triage within 7 days and will correct or remove affected records in a patch release |
| (e.g. v1.0.1) with a CHANGELOG note. |
|
|
| ## Using AppSecBench for disclosure research |
|
|
| - The benchmark is suitable for *evaluating* detection/remediation tools. It is **not** a list of |
| live targets. |
| - If your research surfaces a previously unknown vulnerability class or a flaw in a referenced |
| third-party library, follow that project's own coordinated disclosure process. Do not publish |
| working exploits for unpatched software without prior coordination. |
| - When you publish results that rely on AppSecBench, include the version (`v1.0.0`) and the |
| dataset commit hash so findings are reproducible. |
|
|
| ## Scope boundary |
|
|
| AppSecBench covers application/AI/infra security weaknesses at the *code and configuration* level. |
| It is out of scope for: network exploitation tooling, malware, denial-of-service campaigns against |
| third parties, and any activity prohibited by law or by the target's authorization. |
|
|
