skill-guard

**Judgment:** A strong, security-focused pre-install gate for ClawHub skills that substantially improves safety over direct installs, especially for unvetted or newly updated skills. Recommended for any setup that handles sensitive data or runs third-party skills at scale. **What it does well** - Stages skills in `/tmp` and only moves them into the real skills folder if they pass security checks. - Uses mcp-scan (Invariant/Snyk) to analyze the actual skill content, catching AI-specific issues (prompt injections, hidden instructions, data exfil URLs, toxic flows) that VirusTotal and simple review tools typically miss. - Can block installation automatically when threats are detected and quarantines the skill for manual review. - Provides clear CLI usage (`./scripts/safe-install.sh <skill-slug> [--version] [--force]`) and structured exit codes for automation. **Risks / Limitations** - Not a guarantee of safety: advanced or novel attacks may evade static scanning, and users could develop a false sense of security. - Depends on external tooling (mcp-scan, clawhub CLI, uv, and network access); outages or tool changes may cause failures or reduced protection. - May introduce install latency due to staging and scanning, which could be noticeable in bulk or CI environments. **Recommended scenarios** - Installing any new, low-trust, or frequently updated ClawHub skills, especially on machines with secrets or sensitive files. - Organizational or team environments where security policies require a pre-install scan step. - Automated agent setups and CI pipelines that need a programmatic gate (via exit codes) before promoting skills into production.