security-review

name	security-review
description	Security review checks for shell injection, secrets, file access vulnerabilities, and OWASP top 10. Use when reviewing code changes, PRs, diffs, commits, or editing code that handles user input, auth, or file access.
user-invocable	false
allowed-tools	Read, Grep, Glob, Bash

Security Review Skill

Checks

Shell Injection

• Flag unquoted variables in command substitutions: $(cmd $var) → $(cmd "$var")
• Flag eval usage — suggest alternatives (arrays, command)
• Flag xargs without -0 on untrusted input
• Check for command injection via string interpolation

Secret/Credential Detection

• Hardcoded API keys, tokens, passwords in source
• .env files or credential files staged for commit
• Base64-encoded secrets in config files
• Private keys committed to repo

File Access Patterns

• Path traversal vulnerabilities (../ in user-controlled paths)
• Unsafe temp file creation (use mktemp instead)
• World-readable permissions on sensitive files
• Symlink following attacks

OWASP Top 10 (for web code)

• SQL injection (string concatenation in queries)
• XSS (unescaped output in templates/HTML)
• CSRF (missing token validation)
• Insecure deserialization
• Broken authentication patterns

Output Format

When issues found, report as:

⚠️ SECURITY: [category] in file:line
  Issue: description
  Fix: suggested remediation

Vault Privacy Extension (Pillar 12)

Active only when invoked on a vault note (path matches vault/**/*.md or workspace's vault directory). Especially fires on PreToolUse hook when an Edit operation flips a note's frontmatter public: false → public: true — last-mile gate before private→public flow into mojwang.tech via sync-vault.ts.

Allowlist (public-OK references — do NOT flag)

• "Netflix" as employer reference
• Marvin's LinkedIn-public role: "Engineering Manager", "Live Viewing Experiences" (team)
• Public initiatives Netflix has announced (live streaming launch dates publicly disclosed; SDUI publicly published)
• Marvin's preference: prefer functional descriptors ("live streaming product experiences") over team names — team names change with reorgs
• Public talks Marvin has given (titles, venues, conference names)
• Open-source projects Marvin maintains
• General industry concepts ("server-driven UI", "platform engineering")

Blocklist (flag for review — content might leak)

• Internal team names / codenames / project codes — capitalized acronyms or proper nouns adjacent to "team" / "project" / "initiative" that aren't in the allowlist. Pattern: \b[A-Z]{2,}[a-z]?\s+(team|project|initiative)\b
• Specific reports' or peers' names — first-name or full-name references in performance / capability / personnel context. Cross-reference against vault's known public colleagues; flag any name that isn't.
• Salary / metric leaks:
  • Currency adjacent to budget context: \$\d{2,3}[KMB]\b
  • Percent metrics adjacent to "team", "headcount", "budget", "attrition"
  • Headcount numbers (e.g., "team of 12") if specific to non-public team
• Internal tooling specifics — proprietary system names, internal URL hostnames, internal slack channel names
• Quoted blocks without attribution — long quoted text without source citation suggests potentially-third-party material

Verdict format

⚠️ VAULT-PRIVACY: [category] at line N
  Content: "<excerpt>"
  Issue: <which blocklist pattern matched>
  Fix: <recommend rewording, removal, or generalizing>

If no blocklist matches: report ✓ Vault-privacy: clean with the count of allowlist references found (visibility, not flag).

Blocking when fired by the PreToolUse hook on public: true flip. Marvin must address each finding before the Edit completes.