docs/limitations.md

Limitations

• Not a certified security scanner. This project does not provide penetration testing, compliance, or production safety guarantees.
• Rule-based scoring is a starting point. Regex and keyword heuristics miss nuance, context, and adversarial paraphrase. Results can be false positives or false negatives.
• Human interpretation is required. Treat every CaseScore and aggregate metric as a hint for review, not ground truth.
• Benchmark coverage is incomplete. Six families and a finite seed set cannot represent the full space of agent failures or attacks.
• Manual responses only in v0. Without controlled execution of a target system, variance comes from how faithfully pasted outputs reflect real agent behavior.

Use this artifact to structure fragility discussions and compare runs under the same transparent rules—not to assert absolute safety.