Skill Shield

Skill Shield v0.6.1 — Security Auditor

Scan any skill directory for permissions and dangerous patterns. Get a safety rating before you install.

Usage

Run the scanner on a skill directory:

python3 scripts/scan.py /path/to/skill-directory

SARIF Output (GitHub Code Scanning)

python3 scripts/scan.py /path/to/skill-directory --sarif

Output

The script prints two blocks to stdout:

1. A JSON report (between --- JSON START --- and --- JSON END --- markers)
2. A Markdown report (between --- MD START --- and --- MD END --- markers)

Save reports to files

python3 scripts/scan.py /path/to/skill-directory --output-dir /path/to/output

This creates report.json and report.md in the output directory.

Safety Ratings

Detection Capabilities (65 patterns, 11 categories)

• File deletion: rm -rf, shred, unlink, rmtree, rimraf, del /f (7 patterns)
• Network exfiltration: curl POST, wget --post, requests.post, fetch POST, netcat reverse shell, DNS exfil, pipe to curl, socat (9 patterns)
• Environment variable access: process.env, os.environ, .env files, printenv (5 patterns)
• Secret/key access: .ssh/, .gnupg/, private keys, wallets, tokens, passwords, keychain, cloud credentials (8 patterns)
• Privilege escalation: sudo, su, chmod 777, chown, setuid/setgid, doas (6 patterns)
• Code execution: eval, exec(), Function(), child_process, subprocess, os.system, os.popen, compile (8 patterns)
• Data collection: /etc/passwd, /etc/shadow, whoami, hostname, ifconfig, /proc/self (6 patterns)
• Persistence: crontab, systemd, rc.local, shell profile modification, autostart (5 patterns)
• Obfuscation: long base64 strings, hex escapes, charCode, base64 decode, string reversal (5 patterns)
• Cryptocurrency/mining: xmrig/minerd, mining pool URLs, wallet addresses (3 patterns)
• Shell injection: backtick execution, pipe to shell, download-and-execute (3 patterns)

Key Features

Permission Declaration Audit (unique to skill-shield)

Compares tools declared in SKILL.md against tools actually used in code. Reports:

• Undeclared permissions with sensitivity scoring (1-5)
• Unused declared permissions
• Declaration coverage ratio
• Per-tool risk recommendations

Anti-Obfuscation Analysis

Automatically decodes base64 and hex-encoded content, then re-scans decoded output for dangerous patterns. Obfuscated findings receive elevated severity.

Context-Aware False Positive Reduction

• Comments and docstrings: severity reduced by 2
• Markdown code blocks in SKILL.md: severity reduced by 2 (examples, not real code)
• Pattern definition lines in scanner source: skipped entirely
• Original vs adjusted severity shown in reports (e.g., "Low (2←4)")

CWE References

Every detection pattern includes a CWE (Common Weakness Enumeration) reference for professional vulnerability classification.

v0.6.0 — Batch Scan Mode

Batch Scanning

Scan all skills in a directory at once with --batch:

python3 scripts/scan.py /path/to/skills/ --batch
python3 scripts/scan.py /path/to/skills/ --batch --json-summary
python3 scripts/scan.py /path/to/skills/ --batch --json-summary -o /path/to/output

Performance

• Skips venv/node_modules/dist/.git directories automatically
• Caps at 200 script files per skill for safety
• 164 skills scanned in ~8 seconds

Output

• Markdown table with summary stats and per-skill ratings (default)
• JSON summary with --json-summary flag for machine consumption
• Writes batch-summary.json when using -o

v0.5.0 — SARIF Output Format

GitHub Code Scanning Integration

Use --sarif flag to output SARIF 2.1.0 format, compatible with:

• GitHub Code Scanning (upload-sarif action)
• VS Code SARIF Viewer extension
• SARIF Web Viewer

python3 scripts/scan.py /path/to/skill --sarif > report.sarif
python3 scripts/scan.py /path/to/skill --sarif -o /path/to/output

SARIF Features

• Full rule definitions for all 65 detection patterns
• CWE taxonomy references (MITRE CWE 4.14)
• Partial fingerprints for deduplication across runs
• Security severity scores (2.0-10.0 scale)
• Skill-shield metadata in run properties (ratings, recommendation)

v0.4.0 — Security Tool False Positive Fix

String Literal Context Detection

Regex patterns and string constants inside security tools (scanners, auditors) are no longer flagged as dangerous code. The scanner now recognizes when a pattern like rm -rf or curl POST appears inside a string literal (quotes, regex, array) and reduces severity accordingly.

Ignore-Next-Line Support

Add # skill-shield: ignore-next-line above any line to suppress the next finding. Useful for known-safe patterns in security tools.

Results

• 5 security audit tools reclassified from F to A/C/D (agents-skill-security-audit, ai-skill-scanner, skulk-skill-scanner, aoi-prompt-injection-sentinel, aoi-sandbox-shield-lite)
• 0 regressions on known-safe skills

v0.3.0 — Dual Rating System

Security Rating (code safety)

Based purely on dangerous code patterns found in executable files. Not affected by permission declarations.

Compliance Rating (documentation quality)

Based on permission declaration completeness: does SKILL.md declare the tools actually used in code?

Recommendation

Combines both ratings into an actionable recommendation:

• install — Security A/B + Compliance A/B
• install_with_review — Code is safe but permissions undeclared (likely poor docs, not malicious)
• review_required — Security patterns flagged
• do_not_install — Significant security concerns
• documentation_only — No executable code (pure SKILL.md guidance)

False Positive Fixes (v0.3.0)

• JS template literals (backticks) no longer flagged as shell execution
• Variable names (hostname, whoami) no longer flagged as commands
• Shell script normal $VAR usage no longer floods findings
• os.environ.get("KEY") / process.env.KEY reduced severity (standard practice)
• --disable-setuid-sandbox (browser flag) reduced severity
• Documentation-only skills marked as N/A instead of getting A rating

Exit Codes

Support

Tips welcome: 0x6c730bDcfC762e23cE53aD991B75ab9852e87806 (Base) Moltbook: https://www.moltbook.com/u/Yuqian Twitter: @Yuqian0202