Skill scanner

Skill kpatryk/skills/skills/skill-scanner

AI skills

Install

npx -y skills add kpatryk/skills --skill skill-scanner

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Comprehensive guide for Skill Scanner (Cisco AI Defense) to detect prompt injection, data exfiltration, command abuse, and malicious patterns in AI agent skills. Use this skill whenever a user asks to scan `SKILL.md` packages, secure a `skills/` repository, integrate agent-skill security checks in CI/CD (especially GitHub Actions SARIF uploads), tune scanner policies, reduce false positives, add pre-commit scanning, or troubleshoot `skill-scanner` flags and analyzer behavior. Always use this skill for requests involving "skill scanner", "agent skill security", "scan-all", policy presets (strict/balanced/permissive), overlap checks, or threat findings triage.

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

9.1 KB, as published. Nobody here has run it

skill-scanner

skill-scanner is a best-effort security scanner for Agent Skills. It combines static signatures, YARA, pipeline taint analysis, optional behavioral dataflow analysis, LLM semantic analysis, and optional cloud analyzers to detect probable threats in skill packages.

Documentation URLs used to create this skill

Safety and expectation setting

Treat results as best-effort detection, not proof of safety.
"No findings" means no known patterns were detected; it does not guarantee zero risk.
Human review remains required for high-risk skills and production policies.

Install and upgrade

# One-off package install
uv tool install cisco-ai-skill-scanner

# Keep current installation up to date
uv tool upgrade cisco-ai-skill-scanner

# Verify
skill-scanner --help

Environment variables (optional, feature-dependent)

# LLM analyzer / meta analyzer
export SKILL_SCANNER_LLM_API_KEY="..."
export SKILL_SCANNER_LLM_MODEL="anthropic/claude-sonnet-4-20250514"

# Optional cloud analyzers
export VIRUSTOTAL_API_KEY="..."
export AI_DEFENSE_API_KEY="..."

Core commands

# Interactive wizard
skill-scanner

# Single skill directory
skill-scanner scan skills/my-skill

# Whole repo skill tree
skill-scanner scan-all skills --recursive --check-overlap

# CI-friendly SARIF output
skill-scanner scan-all skills \
  --recursive \
  --check-overlap \
  --format sarif \
  --output skill-scanner-results.sarif

Analyzer selection strategy

Analyzer	Enable with	Use when	Notes
Static	default	Always	YAML + YARA signature coverage
Bytecode	default	Python skill packages	Checks `.pyc` integrity
Pipeline	default	Any shell tooling	Taint analysis for command chains
Behavioral	`--use-behavioral`	Python-heavy logic	AST dataflow analysis
LLM	`--use-llm`	Deeper semantic review	Requires API key
Meta	`--enable-meta`	Reducing noisy LLM findings	Pair with `--use-llm`
VirusTotal	`--use-virustotal`	Binary artifacts present	API key + possible upload policy
AI Defense	`--use-aidefense`	Centralized cloud checks	API key required
Trigger specificity	`--use-trigger`	Vague over-broad descriptions	Useful for skill quality checks

Policy and risk tuning

Use policy presets first, then custom YAML if needed:

# Presets
skill-scanner scan-all skills --policy balanced
skill-scanner scan-all skills --policy strict
skill-scanner scan-all skills --policy permissive

# Generate policy file then customize
skill-scanner generate-policy --preset balanced -o skill-scanner-policy.yaml
skill-scanner scan-all skills --policy skill-scanner-policy.yaml

# Interactive policy editor
skill-scanner configure-policy -o skill-scanner-policy.yaml

Important merge behavior for custom policies

Custom policy files merge on top of defaults.
Scalar fields override directly.
Lists replace default lists (they do not append automatically).

When reducing false positives, prefer:

Rule scoping and allowlists in policy.
Severity overrides.
disabled_rules only as a last resort.

Output formats and CI behavior

# Summary for local quick checks
skill-scanner scan-all skills --format summary

# JSON for downstream tooling
skill-scanner scan-all skills --format json --output results.json

# SARIF for GitHub code scanning
skill-scanner scan-all skills --format sarif --output results.sarif

# Markdown report
skill-scanner scan-all skills --format markdown --detailed --output report.md

# Build gate
skill-scanner scan-all skills --fail-on-severity high

Common usage workflows

1) Local repository triage

skill-scanner scan-all skills \
  --recursive \
  --check-overlap \
  --use-behavioral \
  --policy balanced \
  --format table

2) Deep analysis run (high coverage)

skill-scanner scan-all skills \
  --recursive \
  --check-overlap \
  --use-behavioral \
  --use-llm \
  --enable-meta \
  --llm-consensus-runs 3 \
  --policy strict \
  --format markdown \
  --detailed \
  --output skill-security-report.md

3) Lenient parsing for malformed skill packages

skill-scanner scan-all skills --recursive --lenient

pre-commit integration

# .pre-commit-config.yaml
repos:
  - repo: https://github.com/cisco-ai-defense/skill-scanner
    rev: v1.0.0  # use latest release tag
    hooks:
      - id: skill-scanner

Alternative installer:

skill-scanner-pre-commit install

GitHub Actions integration pattern

Use a reusable script in scripts/scan-skills.sh and upload SARIF:

name: Scan Skills

on:
  pull_request:
    paths:
      - "skills/**"
      - "scripts/scan-skills.sh"
      - ".github/workflows/scan-skills.yml"

jobs:
  scan:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      security-events: write
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"
      - uses: astral-sh/setup-uv@v6
        with:
          enable-cache: true
      - run: ./scripts/scan-skills.sh
      - if: always()
        uses: github/codeql-action/upload-sarif@v3
        with:
          sarif_file: skill-scanner-results.sarif

Error handling and troubleshooting

Missing command: install/upgrade via uv tool install|upgrade cisco-ai-skill-scanner.
LLM errors: verify API keys and SKILL_SCANNER_LLM_MODEL.
Unexpected false positives: tune policy (rule_scoping, allowlists, severity overrides) before disabling rules.
Large trees: use file limits in policy, then rerun with custom YAML.
Malformed skills: use --lenient for migration scenarios.

Usage examples (input -> expected output)

Example 1: Basic repository scan

Input: "Scan skills/ and show whether high-risk findings exist."

Command:

skill-scanner scan-all skills --recursive --fail-on-severity high

Expected output:
- Summary per skill.
- Exit non-zero if findings at/above high exist.

Example 2: GitHub code scanning integration

Input: "Produce SARIF for PR annotations."

Command:

skill-scanner scan-all skills --recursive --format sarif --output results.sarif

Expected output:
- results.sarif file suitable for github/codeql-action/upload-sarif.

Example 3: High-noise repo tuning

Input: "Find real issues but reduce documentation false positives."

Command:

skill-scanner scan-all skills \
  --recursive \
  --policy balanced \
  --use-llm \
  --enable-meta

Expected output:
- Lower-noise findings due to policy defaults + meta filtering.

Example 4: Skill-package quality check

Input: "Detect copied/overlapping skill descriptions in the same repo."

Command:

skill-scanner scan-all skills --recursive --check-overlap

Expected output:
- Overlap-related findings where descriptions are suspiciously similar.

Python SDK quick start

from skill_scanner import SkillScanner
from skill_scanner.core.analyzers import BehavioralAnalyzer

scanner = SkillScanner(analyzers=[BehavioralAnalyzer()])
result = scanner.scan_skill("skills/my-skill")
print(result.max_severity, len(result.findings))

Practical execution checklist

When asked to run skill-scanner in a repository:

Identify target (scan for one skill, scan-all for folder trees).
Use --recursive for multi-directory repositories.
Emit SARIF for CI integrations.
Apply policy preset (balanced default, strict for audits).
Add --use-behavioral for Python-heavy skills.
Add --use-llm --enable-meta only when API keys are available and deeper semantic checks are needed.
Explain that findings are risk indicators, not definitive proof of compromise.

One thing to look at

What its author says it does

SKILL.md

skill-scanner

Documentation URLs used to create this skill

Safety and expectation setting

Install and upgrade

Environment variables (optional, feature-dependent)

Core commands

Analyzer selection strategy

Policy and risk tuning

Important merge behavior for custom policies

Output formats and CI behavior

Common usage workflows

1) Local repository triage

2) Deep analysis run (high coverage)

3) Lenient parsing for malformed skill packages

pre-commit integration

GitHub Actions integration pattern

Error handling and troubleshooting

Usage examples (input -> expected output)

Example 1: Basic repository scan

Example 2: GitHub code scanning integration

Example 3: High-noise repo tuning

Example 4: Skill-package quality check

Python SDK quick start

Practical execution checklist

Keep looking