AI Skill Scanner

**Quick judgment:** A useful and generally safe security utility that improves the safety of installing OpenBot/Clawdbot skills, as long as users treat it as a *screening tool* rather than a guarantee of safety. **What it does well** - Statically scans skill folders for: - Credential exfiltration patterns (e.g., sending secrets out, suspicious logging). - Suspicious network calls (e.g., unknown or non-whitelisted domains, unusual HTTP usage). - Obfuscated or encoded code fragments that may hide malicious behavior. - Possible prompt-injection or prompt-manipulation content. - Provides clear severity levels (CLEAN, INFO, REVIEW, SUSPICIOUS, DANGEROUS) and exit codes suitable for automation. - Offers verbose and JSON modes, making it practical both for human review and scripted pipelines. **Risks & limitations** - **Pattern-based only**: Can miss well-crafted or novel obfuscation, and cannot see runtime behavior. - **False positives**: Legitimate tools that access files, network, or do encoding/decoding may be flagged as REVIEW/SUSPICIOUS. - **False sense of security**: Passing the scan does **not** prove a skill is safe; manual review is still required for higher-severity or complex skills. **Recommended use scenarios** - As a **mandatory pre-install scan** for any third-party or untrusted skill from ClawHub or other sources. - As a **CI/CD or approval gate**: fail or require security review when exit code indicates REVIEW/SUSPICIOUS/DANGEROUS. - As a **triage tool** to prioritize which skills need deeper manual security review. Overall: Strongly recommended as one layer in a defense-in-depth workflow for managing skill security, but it should always be combined with human review and broader security practices.