skillspector trial.stub

---

id: skillspector-trial name: Skillspector repo_path: ~/Documents/sw30labs/repos/skillspector-trial remote_url: https://github.com/sw30labs/skillspector-trial.git primary_language: JavaScript framework: none (vanilla JS, single-file) ingested: 2026-07-04 last_commit_date: 2026-07-03 category: ai-security stacks: [agentic]

Skillspector

Purpose

A single-file, offline web app that scans Claude / Agent Skills for security and quality problems. Drag in a skill — a folder, a .zip, or a .skill file — and get a graded report (A–F) with findings, evidence excerpts, and a capability breakdown. Everything runs client-side; nothing is uploaded and there are no network calls.

Architecture

• index.html — the built, self-contained app (the only file a user opens; works from file://).
• src/engine.js — the scan engine: pure-JS zip reader, rule catalog, scoring. Also runs under Node ≥18, which is how downstream tooling consumes it.
• src/ui.js, src/style.css, src/template.html — drop zone, scan animation, report rendering, dark-first theme.
• build.mjs — inlines src/* into the single index.html.

Rules key on the shape of an attack, not on capability: a newsletter skill that sends mail via SMTP with an env-var password grades A; a skill that reads ~/.aws/credentials and POSTs it to a webhook grades F. Documentation-aware (a SKILL.md warning against rm -rf / isn't punished) and placeholder-secret aware.

Rule catalog

• Security SEC-001…010: prompt-injection phrasing, hidden/invisible-unicode + ASCII smuggling, data exfiltration, dangerous shell, dynamic-code/obfuscation, hardcoded secrets, sensitive-path access, persistence/env tampering, safety-bypass instructions, remote fetch-and-run.
• Quality QUA-001…011: SKILL.md presence/body, frontmatter validity, name/description quality, trigger guidance, file length, broken references, junk files, bundle size, unreferenced scripts, binary blobs.
• Scoring: start 100, deduct per finding (critical −30, high −15, medium −7, low −3), grade A–F. Stable rule IDs are a contract.

CI/CD

Node test suite (tests/run-tests.mjs, 74 tests, zero dependencies). Rule IDs and severities are a stable API documented in SPEC.md.

Note

Name-collides with NVIDIA's SkillSpector (a separate agent-skill scanner) — worth disambiguating before wider publication.

Tags

security llm pentesting

Cross-references

• oscal-skills-guardrails