Skill Security Scan

${var} — Path to a SKILL.md file or skill directory to scan. If empty, scans all skills in skills/.

If ${var} is set, scan only that path. Otherwise, scan all skill directories.

Today is ${today}. Your task is to audit skill files for security vulnerabilities before they can be executed.

Threat Model

Imported skills are markdown files that instruct Claude Code to take actions. A malicious skill could:

• Shell injection: Execute arbitrary commands via unquoted variables, eval, backticks, or $(...) in bash blocks
• Secret exfiltration: Send environment variables, tokens, or file contents to external URLs via curl/wget/fetch
• Path traversal: Access files outside the repo using ../ or absolute paths
• Prompt injection: Override CLAUDE.md safety rules with embedded instructions ("ignore previous instructions", "you are now...")
• Destructive commands: Run rm -rf, git push --force, or other irreversible operations

Steps

1. Determine scan scope:

   • If ${var} is a file path, scan that file only.
   • If ${var} is a skill name, scan skills/${var}/SKILL.md.
   • If empty, scan all skills/*/SKILL.md files.

2. Run the scanner on each skill file using ./skills/skill-security-scan/scan.sh:

   bashCopy

   ./skills/skill-security-scan/scan.sh <path-to-SKILL.md>
   

   The scanner checks for the threat categories above and outputs findings with severity levels (HIGH, MEDIUM, LOW).

3. Check trusted sources: Read skills/security/trusted-sources.txt. Skills from trusted sources get a reduced scan (format validation only, skip content analysis). The source is determined by checking git remote or the skill's frontmatter for an origin field.

4. Generate report: For each scanned skill, produce:

   Copy

   [PASS/WARN/FAIL] skill-name
     HIGH: description (if any)
     MEDIUM: description (if any)
     LOW: description (if any)
   

   • FAIL = any HIGH severity finding
   • WARN = MEDIUM findings only
   • PASS = no findings or LOW only

5. Save report to articles/security-scan-${today}.md with full details.

6. Notify via ./notify if any skills FAIL:

   Copy

   *Security Scan — ${today}*
   Scanned N skills: X passed, Y warnings, Z failed.
   Failed: skill1 (reason), skill2 (reason)
   

   If all pass, log "SECURITY_SCAN_OK" and skip notification.

7. Log results to memory/logs/${today}.md.