skill-security-scan（技能安全扫描）

Skill Security

Run these checks before installing ANY skill, package, or plugin. Always warn the user before proceeding.

Pre-Install Checklist

1. Source Verification

• Is it from clawhub.com (vetted) or a random npm package?
• Who's the author? Do they have other packages / reputation?
• Is there a GitHub repo? Check for recent commits, open issues, maintainer activity.

2. Popularity Check

• npm info <package> — check weekly downloads, last publish date, version history
• Low downloads (< 100/week) + recent publish = higher risk

3. Dependency Audit

• npm info <package> dependencies — how many deps does it pull in?
• More dependencies = larger attack surface
• Flag packages with 50+ transitive dependencies

4. Lifecycle Scripts (HIGH RISK)

• Check for preinstall, install, postinstall scripts — these run arbitrary code
• npm info <package> scripts or inspect package.json
• pnpm blocks these by default; npm does NOT
• If lifecycle scripts exist, flag it explicitly to the user

5. Scan After Install

• npm audit after install to catch known vulnerabilities

6. Check for Dynamic Content

• Search the skill code for URLs that are fetched at runtime
• Skills that download and execute content from external endpoints can change behavior after install

Red Flags — Stop and Ask

• 🚩 Package published very recently with no history
• 🚩 Maintainer has no other packages or reputation
• 🚩 Package name similar to a popular one (typosquatting: e.g., reqeust vs request)
• 🚩 Requests permissions beyond what it claims to do
• 🚩 No GitHub repo, or repo is empty/suspicious
• 🚩 Postinstall scripts that download from unknown URLs
• 🚩 Dynamic content fetching — skill calls external URLs at runtime (2.9% of ClawHub skills do this; payload can change after install)
• 🚩 Base64 in install instructions — ClickFix social engineering pattern (fake errors → paste base64 command → malware)
• 🚩 New uploader, bulk uploads — single user uploading many skills rapidly (ClawHavoc: 354 skills from one account)
• 🚩 Skill references credential paths — ~/.openclaw/credentials/, ~/.clawdbot/.env, .env files

Core File Protection

Skills are NEVER allowed to modify these files without explicit user approval:

• SOUL.md — agent identity
• AGENTS.md — agent rules
• IDENTITY.md — agent metadata
• USER.md — user's personal info
• MEMORY.md or memory/*.md — agent memories
• TOOLS.md — infrastructure notes

Data Exfiltration Checks

After installing a skill, before running it:

1. Read SKILL.md — understand what the skill does
2. Check file paths — does it reference paths outside its own directory and workspace?
3. Check exec commands — does it curl to unknown domains?
4. Check write/edit calls — are target paths outside workspace/?
5. Check for credential access — .env, ~/.ssh/, ~/.gnupg/, API keys
6. Check for dynamic content — does it fetch and execute content from external URLs at runtime? (Snyk: 2.9% of ClawHub skills do this)

ClawHub-Specific Checks

When installing from ClawHub:

• Check uploader history — is this a new account? Do they have other skills? Bulk uploads from unknown accounts = red flag (ClawHavoc: 354 skills from one malicious account)
• Check skill stars/reviews — community feedback is a signal, not proof
• Check VirusTotal — OpenClaw has a VirusTotal partnership; check if the skill has been scanned
• Verify semantic versioning — legitimate skills typically have version history and changelogs
• Search for security reports — search web for "[skill-name] malicious" before installing

Known Attack Campaigns (as of March 2026)

Reference for identifying patterns:

• ClawHavoc — 300+ coordinated skills, ClickFix social engineering, downloads Atomic Stealer (AMOS)
• AuthTool — dormant payload, activates on specific natural language prompts, establishes reverse shell
• Hidden Backdoor — fake "Apple Software Update" during install, encrypted tunnel to attacker
• Credential Exfiltration — targets ~/.clawdbot/.env and ~/.openclaw/credentials/ for API keys

Red Flags — Skill Behavior

• 🚩 Skill reads .env or credential files
• 🚩 Skill makes network requests to unfamiliar domains
• 🚩 Skill writes to paths outside workspace
• 🚩 Skill modifies core files (SOUL, AGENTS, MEMORY, USER, IDENTITY, TOOLS)
• 🚩 Skill sends data to external URLs not part of stated purpose

What To Tell The User

Before installing, give a brief summary:

"⚠️ Installing [package]: [downloads/week], [last updated], [dep count] deps, [lifecycle scripts?]. Looks [clean/sketchy] — proceed?"

If red flags found:

"🚩 Flags on [package]: [list issues]. Want me to proceed anyway?"