Skill Scanner

**Judgement:** A generally useful and relatively safe local security-audit tool for Clawdbot/MCP skills, best used as a *first-line* static scanner rather than a definitive security authority. **What it does well** - Scans skill folders for common malicious or risky patterns: data exfiltration, system modification, crypto-mining, backdoors, obfuscation, and arbitrary-code-execution risks. - Runs locally with minimal dependencies (Python stdlib), reducing external data exposure and simplifying deployment. - Provides both CLI and optional Streamlit Web UI, and can output Markdown or JSON—good for both manual review and automation/CI. **Key risks / limitations** - Likely pattern/signature-based static analysis only: it can miss novel, subtle, or heavily obfuscated malware (false negatives). - May over-flag benign but powerful code (e.g., legitimate filesystem or network operations) as suspicious (false positives), requiring human triage. - No evidence (from metadata) of formal testing, rule quality, or coverage guarantees; results should not be treated as a complete security audit. - As a filesystem scanner, it needs read access to code; while that is expected, organizations should still control who can run it and where reports are stored. **Recommended scenarios** - Pre-install and pre-update checks on third-party or untrusted skills, especially in production, corporate, or privacy-sensitive environments. - CI/CD gates and automated policy enforcement to prevent obviously risky skills from being deployed without review. - Periodic security hygiene scans of an existing skill set, combined with manual code review for anything high-impact or heavily flagged. **Overall:** Suitable as a helpful security layer for screening skills, but it should complement—not replace—manual review, broader security tooling, and standard operational safeguards.

**Quick judgment:** Skill Scanner is a focused static-analysis style tool that meaningfully improves security hygiene for Clawdbot/MCP skills by flagging common malicious or risky code behaviors. It is particularly valuable as a pre-installation gate or part of a CI/security review pipeline, but it should be treated as an aid—not a definitive security authority. **Strengths & benefits** - Targets real-world risks: data exfiltration, system modification attempts, backdoors, arbitrary code execution, and crypto-mining indicators. - Low operational friction: Python 3.7+ with standard library only; optional Streamlit UI for less technical users. - Machine- and human-friendly outputs (JSON/Markdown) make it easy to integrate into automated checks and manual review processes. - Designed specifically around agent/skill safety, aligning with platform security needs rather than generic endpoint protection. **Key risks & limitations** - **False negatives:** Static pattern-based scanning can miss sophisticated, novel, or well-obfuscated malware; it cannot guarantee a skill is safe. - **False positives:** Legitimate advanced functionality (e.g., system calls, network access, cryptographic or multiprocessing code) may be flagged as suspicious, requiring human interpretation. - **Over-reliance risk:** Treating a “clean” report as a security guarantee could lead to deploying unsafe skills; results must be combined with code review, provenance checks, and runtime controls. - **Scope limits:** Focuses on the skill’s own code and obvious behaviors; may not fully capture risks introduced by external services, APIs, or dynamically loaded components. **Recommended scenarios** - Use as a **mandatory pre-install screen** for any third-party skill before enabling it in production or shared environments. - Integrate into **CI/CD for skill development** to automatically flag risky coding patterns before publishing. - Employ in **periodic security audits** of existing skill inventories, especially after updates, configuration changes, or incident investigations. - Combine with **manual code review and runtime sandboxing** for high-impact or highly privileged skills to build a layered defense rather than relying solely on automated scanning.