| # Intended Uses |
|
|
| AppSecBench is built for **defensive security** research, evaluation, and education. |
|
|
| ## In scope |
|
|
| - **LLM evaluation** — measuring a model's ability to detect, classify (CWE/OWASP), explain, |
| score severity (CVSS), recommend fixes, and generate secure code for the covered weakness types. |
| - **SAST / scanner benchmarking** — comparing static-analysis, secret-scanning, and |
| IaC-scanning tools on a common, versioned set of cases. |
| - **Secure-coding assistants** — fine-tuning or prompting models for secure code review and |
| remediation (subject to the MIT license and the ethical-use note). |
| - **Academic research** — reproducible experiments in application security, AI security, and |
| secure software engineering. |
| - **Curricula** — teaching secure-coding patterns via paired vulnerable/secure examples. |
|
|
| ## Out of scope / prohibited |
|
|
| - Offensive operations against systems you are not authorized to test. |
| - Training models to *generate* exploits or attacks rather than defenses. |
| - Repackaging the dataset to violate the licenses of referenced standards or to misrepresent its |
| provenance. |
| - Any use that breaks applicable law or the terms of the platforms where models/tools are deployed. |
|
|
| ## Recommended evaluation protocol |
|
|
| 1. Pick a split (`dataset/test.jsonl` for held-out eval). |
| 2. Run your model/tool; collect outputs keyed by `benchmark_id`. |
| 3. Score with `scripts/evaluate.py` (or your own judge using the weighted rubric). |
| 4. Report overall and per-language / per-vulnerability / per-difficulty breakdowns, the dataset |
| version, and the commit hash. |
|
|
| See `docs/methodology.md` and `docs/LIMITATIONS.md`. |
|
|
