Ethical Considerations
AppSecBench intentionally contains vulnerable code. This document explains how the data is constructed and the safeguards around its release.
What the data is
- Synthetic and minimal. Every vulnerable snippet isolates one weakness in a few lines. It is not extracted from, and does not reproduce, any real application, product, or copyrighted source.
- Non-weaponized. Exploit sketches are illustrative proof-of-concept patterns (e.g. a sample malicious URL or payload shape), not turnkey exploits, tooling, or targeting information.
- Educational framing. Each record pairs the vulnerable code with secure code, an explanation, and references to authoritative guidance (OWASP, CWE, ASVS).
Why release it
The security community needs open, reproducible benchmarks to (a) measure and improve LLM secure- code capability, (b) benchmark SAST/secret-scanning/IaC-scanning tools on equal footing, and (c) support defensive academic research. Withholding such data would cede evaluation to opaque, closed benchmarks.
Safeguards
- The dataset is MIT-licensed for defensive use (research, education, tool evaluation, training).
- AI/LLM cases (prompt injection, RAG, MCP, agent security) demonstrate defensive controls, not attack tooling.
- Infrastructure cases show misconfiguration and the corrected, hardened configuration side by side.
- The dataset does not contain working exploits against third parties, credential material
beyond obvious placeholders (
sk_live_9f8a...), or instructions for unauthorized access.
Responsible use
- Do not use the vulnerable snippets to attack systems you are not authorized to test.
- When training models, ensure the secure-code pairs are presented with their explanations so the model learns remediation, not mere vulnerability replication.
- Cite the dataset and link to this document when publishing results.
Contact
Report concerns via the repository's SECURITY.md disclosure process.