Skill Auditor

Skill Auditor v3.1.0

The definitive security scanner for OpenClaw/ClawHub skills. Best-in-class detection across 18 security checks including prompt injection detection — the first scanner to catch agent manipulation attacks in skill documentation. 5-dimension trust scoring, trend tracking, diff analysis, and benchmarking. Zero false positives on legitimate skills.

When to Activate

1. Installing a new skill from ClawHub - run inspect.sh for full pre-install validation
2. Auditing existing skills - use audit.sh to scan any skill directory
3. Generating trust scores - use trust_score.py for 0-100 rating across 5 dimensions
4. Comparing skills - use trust_score.py --compare for side-by-side analysis
5. Tracking improvements - use trust_score.py --save-trend to monitor score over time
6. Reviewing updates - use diff-audit.sh to compare before/after versions
7. Batch scanning - use audit-all.sh or benchmark.sh for fleet-wide analysis

Quick Start

# Audit a single skill
bash audit.sh /path/to/skill

# Trust score (0-100 across 5 dimensions)
python3 trust_score.py /path/to/skill

# Compare two skills side by side
python3 trust_score.py /path/to/skill1 --compare /path/to/skill2

# Track score over time
python3 trust_score.py /path/to/skill --save-trend
python3 trust_score.py /path/to/skill --trend

# Diff audit (before/after update)
bash diff-audit.sh /path/to/old-version /path/to/new-version

# Benchmark against a corpus
bash benchmark.sh /path/to/skills-dir

# Inspect a ClawHub skill before installing
bash inspect.sh skill-slug

# Audit all installed skills
bash audit-all.sh

# Generate a markdown report
bash report.sh

# Run test suite (28 assertions)
bash test.sh

Guardrails / Anti-Patterns

DO:

• ✓ Always audit skills before installing from untrusted sources
• ✓ Review trust scores - reject skills scoring below 60 (D grade)
• ✓ Use diff-audit.sh when updating skills to catch regressions
• ✓ Use --json output for CI/CD pipeline integration
• ✓ Run --save-trend periodically to track skill health

DON'T:

• ✗ Install skills scoring below 40 (F grade) without extensive manual review
• ✗ Ignore CRITICAL findings - they indicate potential security threats
• ✗ Blindly add skills to allowlist without understanding why they access credentials
• ✗ Skip audit because a skill is "popular" or "official"

Security Checks (18 total)

Context-aware: credential mentions in documentation are INFO, not CRITICAL.

Trust Score (5 Dimensions)

Grades: A (90+), B (75+), C (60+), D (40+), F (<40)

Comparative Scoring

python3 trust_score.py /path/to/skill-a --compare /path/to/skill-b

Shows per-dimension deltas and overall score difference.

Trend Tracking

python3 trust_score.py /path/to/skill --save-trend   # Record score
python3 trust_score.py /path/to/skill --trend         # View history

Stores up to 50 entries per skill in trust_trends.json.

Tools

Test Suite

12 test skills (8 malicious, 4 clean) with 28 automated assertions:

bash test.sh

Malicious fixtures: credential harvest, obfuscated payload, sensitive fs reads, crypto wallets, time bombs, symlink attacks, prompt injection, download-execute, privilege escalation. Clean fixtures: basic skill, credential docs (false positive check), network skill, dotfiles skill.

Exit Codes

• 0: PASS / safe to install
• 1: REVIEW / warnings found
• 2: FAIL / critical issues
• 3: Error / bad input

Changelog

See CHANGELOG.md for full version history.