What problem does it solve?
Validating that a security-detection harness actually improves over static or single-pass LLM approaches requires a ground-truth corpus and measurable gates; this Skill runs the upstream @metaharness/darwin security bench to evolve a champion detector and grade it on TPR, FPR, patch-pass, reproduction, and unsafe-output metrics.
Core Features & Use Cases
- Evolutionary Benchmarking: Shells out to
npx @metaharness/darwin security bench with configurable population, cycles, and seed to evolve a champion detection harness against a 10-vulnerability / 9-decoy corpus.
- Gate and Baseline Grading: Parses the markdown report into structured JSON covering 11 pass/fail gates and a four-row baseline table (static-only, LLM single-pass, fixed agent, Darwin champion).
- Nightly ADR-155 Integration: Emits trajectory records for ruflo's self-learning security harness, with
--alert-on-fail exit codes and graceful degradation when the upstream package is absent.
- Use Case: Run the bench nightly in CI to detect drift in detection quality; if the champion's fitness drops week-over-week, investigate mutator policy or corpus changes before trusting downstream learning loops.
Quick Start
Ask the agent to run the security bench with default settings, for example: run the harness security bench with population 2 and cycles 1 and alert me if it fails.
