skill-audit-framework（技能审核框架）

Skill Auditor 🔍

A structured framework that teaches your agent how to audit ClawHub and MCP skills before you install them. Not a scanner — a systematic review methodology.

Unlike automated scanners that give false confidence, Skill Auditor walks through what matters: permissions, behavior, credentials, and persistence — so you understand exactly what a skill will do on your system.

Why this exists

• 13.4% of ClawHub skills have critical security issues (Snyk ToxicSkills study)
• 341 malicious skills were found in a single campaign (ClawHavoc incident, Feb 2026)
• Automated scanners can miss context-dependent threats and provide false security
• Understanding what you're installing is better than trusting a green checkmark

How to use

Ask your agent to audit any skill before installing:

Audit this skill before I install it: [skill-name or URL]

Review the security of @author/skill-name on ClawHub

I want to install [skill]. Is it safe?

Audit Framework

The agent follows a 6-domain checklist. Each domain produces a PASS / WARN / FAIL verdict.

1. Identity & Provenance

• [ ] Author has a GitHub profile with other projects
• [ ] Skill has a public source repository (not ClawHub-only)
• [ ] Repository has commit history (not a single-commit dump)
• [ ] Author identity is consistent across platforms
• FAIL if: No source repo, no author history, single-commit repo

2. Permission & Scope Analysis

• [ ] requires.env only lists credentials the skill actually uses
• [ ] No credentials unrelated to the skill's purpose
• [ ] File access limited to workspace directory
• [ ] No requests for system-wide permissions
• FAIL if: Requests credentials beyond stated purpose, accesses files outside workspace

3. Behavior vs Description Match

• [ ] Every file in the skill serves the stated purpose
• [ ] No network calls to undeclared endpoints
• [ ] No data exfiltration patterns (sending user data to external URLs)
• [ ] Script behavior matches what SKILL.md describes
• FAIL if: Hidden functionality, undeclared network calls, description mismatch

4. Credential & Secret Handling

• [ ] API keys stored in env vars, not hardcoded
• [ ] No credentials logged or written to non-protected files
• [ ] OAuth tokens have minimal required scopes
• [ ] Cached tokens stored in workspace, not system-wide
• FAIL if: Hardcoded secrets, credentials in logs, excessive OAuth scopes

5. Persistence & Side Effects

• [ ] Files written only within workspace boundaries
• [ ] No system-level modifications (crontab, /etc/, systemd)
• [ ] No auto-start or background processes installed
• [ ] Uninstall is clean (no orphaned files or processes)
• FAIL if: System modifications, persistent background processes, dirty uninstall

6. Dependency & Supply Chain

• [ ] Dependencies are well-known packages (not obscure single-author libs)
• [ ] No curl | bash or curl | python install patterns
• [ ] No post-install scripts that download additional code
• [ ] Package versions are pinned (not latest)
• FAIL if: Unknown dependencies, pipe-to-shell installs, unpinned versions

Output Format

The agent produces a structured report:

## Skill Audit Report: [skill-name]

Author: [name] | Source: [repo URL or "ClawHub only"]
Version: [X.Y.Z] | Files: [count] | Scripts: [count]

### Verdicts

| Domain                    | Verdict | Notes                |
|---------------------------|---------|----------------------|
| Identity & Provenance     | PASS    |                      |
| Permission & Scope        | WARN    | Requests broad perms |
| Behavior vs Description   | PASS    |                      |
| Credential Handling       | PASS    |                      |
| Persistence & Side Effects| FAIL    | Writes to /etc/      |
| Dependency & Supply Chain | PASS    |                      |

### Overall: ⚠️ WARN — Review flagged items before installing

### Flagged Items
1. [Domain]: [Specific issue and recommendation]

### What to Ask the Author
1. Why does the skill need [permission X]?
2. Can [flagged behavior] be made opt-in?

Limitations

• This is a review framework, not a deterministic scanner
• The agent reads and reasons about skill files — it cannot execute or sandbox them
• Always read the source code yourself for high-privilege skills
• A PASS verdict means no issues were found, not that the skill is guaranteed safe

Trust Hierarchy

When evaluating skill trust, consider this hierarchy:

1. Highest trust: Open-source on GitHub + active maintainer + ClawHub Benign scan + you read the code
2. Moderate trust: GitHub repo exists + ClawHub Benign scan + reasonable permissions
3. Low trust: ClawHub-only (no source repo) + Suspicious scan + broad permissions
4. No trust: No source, no author history, requests unrelated credentials