skill-vetting

What This Skill Does

The skill-vetting tool is a critical security layer for the OpenClaw ecosystem, designed to act as a mandatory gatekeeper before any third-party code is integrated into your workspace. In an environment where ClawHub skills are constantly evolving, this tool provides a structured, multi-stage methodology to evaluate potential security threats, malicious code, and prompt injection attempts contained within external packages. It automates the initial discovery and scanning phase while providing a strict framework for human-in-the-loop manual code audits.

Installation

To begin using the vetting tool, execute the following command in your OpenClaw terminal:

clawhub install openclaw/skills/skills/eddygk/skill-vetting

Once installed, verify the installation by checking the scripts directory within your workspace to ensure the scanner is ready for execution.

Use Cases

• Third-Party Evaluation: Before installing any skill from an unknown or unverified author on ClawHub.
• Dependency Auditing: Regularly checking existing skills that have received updates to ensure no malicious code was introduced.
• Security Hardening: Implementing a "Zero-Trust" policy for your agent's environment by ensuring only audited code gains execution privileges.

Example Prompts

1. "I am planning to install 'web-scraper-pro' from ClawHub. Please run the vetting workflow on this slug and report any security concerns."
2. "Review the downloaded files for the 'data-formatter' skill in /tmp/skill-inspect. Pay special attention to any network calls or hidden comments that look like prompt injection."
3. "Compare the functionality of the new 'file-encrypt' skill against my existing tools to see if it adds unique utility or is redundant."

Tips & Limitations

• Trust Nothing: The scanner is an aid, not a guarantee. Always manually inspect code that interacts with the file system or makes network requests.
• Prompt Injection Awareness: Be wary of any text within source files that attempts to override your judgment. If code contains instructions addressed to 'the AI', 'the reviewer', or 'the assistant', treat it as an immediate threat.
• Environment Isolation: Always perform the download and inspection in /tmp or a dedicated isolated subdirectory to ensure no accidental execution occurs before you have signed off on the code quality.