Security Audit

Security Audit

Audit the code at $ARGUMENTS against production-quality security standards.

If no path is provided, audit the most recently changed files: !git diff --name-only HEAD~1 2>/dev/null | head -30

Scope

• Single file → audit that file deeply
• Directory → recurse, prioritize entry points (routes, handlers, IaC, CI configs, Dockerfiles)
• PR diff → audit only changed lines + their surrounding context
• "the whole repo" → full audit (warn about runtime; suggest tooling)

Audit Categories

For each piece of code, run the categories below. Only report actual issues. Calibrate severity by exploitability + blast radius (see [threat-modeling.md](../bridgesecurity/references/threat-modeling.md)).

1. Injection (CWE-79, CWE-89, CWE-78, CWE-94, CWE-77)

SQL / NoSQL / OS command / code / template / expression-language / LDAP / XPath. Look for: string concat into queries, eval, child_process.exec, dangerouslySetInnerHTML, pickle.loads, ORM raw/literal/Unsafe APIs.

2. Broken Access Control (CWE-862, CWE-863, CWE-639, CWE-352)

IDOR, BOLA, BOPLA, missing auth on routes, mass assignment, missing CSRF, client-only checks, Server Actions without auth() call. Trace every state-changing handler — what's the first non-trivial line?

3. Cryptographic Failures (CWE-327, CWE-328, CWE-330, CWE-916)

Weak hashes (MD5, SHA1, bcrypt , curl … | sh in setup scripts, postinstall scripts, vulnerable deps from npm audit`.

14. Open Redirect (CWE-601)

router.push(searchParams.get('next')) without allowlist; bypasses like //evil.com.

15. ReDoS (CWE-1333)

Catastrophic backtracking patterns: (a+)+, (a*)*, (a|aa)+.

16. Server Action / RPC Authorization

Every 'use server' function — first line auth check + ownership verification + input schema?

17. Sensitive Data Exposure (CWE-200)

Stack traces in responses, full user objects hydrated to client, secrets in logs, error messages with file paths.

18. CSRF / CORS / Cookies (CWE-352)

Missing CSRF token, cors({ origin: '*', credentials: true }), cookies without httpOnly / Secure / SameSite.

19. Markdown / Output Rendering Hazards

Image URL constructed from secrets / untrusted content (exfil channel — see BridgeWard).

20. LLM-Specific (when applicable)

LLM05 (output passed to dangerous sink), LLM06 (excessive agency), LLM07 (secrets in system prompt).

Output Format

# BridgeSecurity Audit Report

**Target**: 
**Scanned**: 
**Date**: 

## Summary

| Severity | Count |
|---|---|
| Critical | N |
| High | N |
| Medium | N |
| Low | N |
| Info | N |

## Findings

### [Critical] 
- **File**: `path/to/file.ts:42`
- **Class**: 
- **CWE**: CWE-89 (SQL Injection)
- **OWASP**: A03 Injection
- **Code**:
  ```ts
  const result = await db.query(`SELECT * FROM users WHERE email = '${email}'`);
  ```
- **Why it's a problem**: 
- **Fix**:
  ```ts
  const result = await db.query('SELECT * FROM users WHERE email = $1', [email]);
  ```
- **References**: [CWE-89](https://cwe.mitre.org/data/definitions/89.html), [BridgeSecurity language-patterns](../bridgesecurity/references/language-patterns.md#sql--orm-injection)

[next finding...]

## Conclusion

Severity Levels

• Critical: direct RCE / full data exposure / auth bypass / mass exfiltration. Fix before merge.
• High: significant data exposure / privilege escalation / DoS / IDOR. Fix this sprint.
• Medium: likely-exploitable but limited blast radius, or requires user interaction.
• Low: defense-in-depth gap.
• Info: best-practice deviation worth noting.

Calibrate to production impact, not theoretical purity. A Math.random() for an unguessable element ID is Info. The same code generating a session token is Critical.

Important Rules for Auditor

1. Read-only. Suggest fixes; don't apply them unless explicitly asked.
2. Quote, don't paraphrase. Show the exact vulnerable code.
3. Concrete fixes. Show the patched code, not just "use parameterized queries".
4. Map to CWE + OWASP. Use canonical IDs.
5. No false positives if avoidable. Verify before flagging:

• Is the input actually user-controllable, or constant?
• Is this a test fixture / example?
• Is there sanitization on the way to the sink?

1. Calibrate severity by reachability. Unreachable vuln in dead code = Info.
2. Don't dump tool output. If you ran semgrep / trivy / gitleaks, synthesize. Deduplicate. Prioritize.
3. Respect environment files. Never read or display .env* contents (project rule). You can flag references to env files (e.g., a Dockerfile that copies .env).

Default behavior when $ARGUMENTS is empty

Audit the most recently changed files (last commit / current diff). If outside a git repo, ask the user:

> "What would you like me to audit? Options: (1) a file or directory path, (2) git diff HEAD~1, (3) the whole repo (slower), (4) a specific concern (e.g., 'all server actions', 'all SQL queries', 'all Dockerfiles')."

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: bridge-mind
• Source: bridge-mind/BridgeSecurity
• License: MIT
• Homepage: https://www.bridgemind.ai

Install and usage instructions live in the source repository linked above.