# Known Limitations

AppSecBench is a strong but *bounded* benchmark. Be aware of these limitations when interpreting
results.

## Coverage & realism

- **Synthetic, isolated snippets.** Each case demonstrates one weakness in a few lines. Real code
  interleaves multiple concerns; high scores here do not guarantee strength on large, messy
  codebases.
- **Not exhaustive.** 34 vulnerability classes are covered; many CWE entries and emerging weakness
  types (e.g. some supply-chain, post-quantum, or domain-specific issues) are not yet represented.
- **Framework breadth is curated.** Only the frameworks in the supported list appear; other
  popular stacks (e.g. Rails, Flask alternatives) are not covered in v1.0.0.
- **AI/LLM cases are conceptual.** Prompt-injection / RAG / MCP / agent examples are illustrative
  of the *control* (isolation, scoping, authz, sandboxing), not a full agent harness.

## Tooling & checks

- **Heuristic syntax checks for some languages.** Rust, C#, Java, Kotlin, Swift, PHP, C/C++ are
  validated with delimiter-balance heuristics when their compiler is absent; a "PASS" on those is
  a balance check, not a guarantee of compilability. (Python/JS/TS/Go/YAML/Bash use real toolchains
  when installed.)
- **CVSS reflects the archetype, not a specific deployment.** Scores use dominant realistic
  exploitability/impact vectors; a given app's actual risk may differ.

## Evaluation

- **The reference grader is transparent but shallow.** `scripts/evaluate.py` uses keyword/format
  matching. For publication-grade results, pair it with an LLM judge or human review; the rubric
  weights are provided for that purpose.
- **Severity priors are estimates.** `expected_false_positive_probability` /
  `expected_false_negative_probability` encode difficulty-based priors, not empirically measured
  rates.

## Suggested improvements for v1.1+

- Grow to 1,000+ records with broader language/framework coverage (Rails, Vue, Angular, Scala).
- Add multi-file / multi-function cases and end-to-end mini-apps.
- Add an LLM-judge harness and inter-rater reliability on a labeled sample.
- Add executable test oracles (do the secure snippets actually pass a security test?).
- Expand AI/LLM coverage with full agent/MCP server fixtures.
- Add license/PII redaction checks across more ecosystems.

See `CHANGELOG.md` for the version plan.