How to do a security audit of a OpenAI Codex app?

Why a OpenAI Codex-specific audit (not a generic web audit)

A generic OWASP audit will tell you your OpenAI Codex app "needs CSP headers." A OpenAI Codex-aware audit tells you that your specific OpenAI Codex app has an RPC function callable without auth or a service key in a client bundle — the issues that actually appear when OpenAI Codex apps get compromised. The difference in output value is why the audit should be scoped to OpenAI Codex's real failure modes.

Step 1 — Fingerprint the deployment

Confirm the OpenAI Codex stack components: database (supabase, firebase, postgres), hosting, auth provider, third-party integrations. For OpenAI Codex apps this is often visible in the Supabase endpoint URL in network requests. Document every component — each is an independent audit target.

Step 2 — Automated scan with OpenAI Codex-aware rules

Run VAS against the deployed URL. The scan probes the specific issue classes found in OpenAI Codex apps: secrets detection, input validation, auth security, data authorization. This is the 80/20 — most critical and high findings surface here. Fix anything critical before continuing to manual steps.

Step 3 — Manual Row Level Security (RLS) policies review

Open the Supabase dashboard → Authentication → Policies. For each table: is RLS enabled? Do policies check `(select auth.uid()) = user_id` or equivalent? Are there policies scoped to the anon role that shouldn't exist? The automated scan catches missing RLS; this step catches overly permissive RLS — a subtler but equally dangerous failure mode.

Step 4 — Authentication & authorization probing

Test every endpoint with no session (expect 401), with a valid session for a different user (expect 403 on user-owned resources), and with session tokens that have been tampered with (expect 401 if signatures are enforced). For OpenAI Codex specifically, watch for weak auth defaults — authentication code may work but lack rate limiting, email verification, or csrf protection.

Step 5 — Re-scan to verify

Fix findings in severity order (critical → high → medium → low), re-scan after each batch of fixes. "I applied the fix" is not evidence — the fix might not have been deployed, might have been partial, or might have been reverted. Only the scan output proves the gap is closed. Log each finding + fix + verification scan for compliance records.

OpenAI Codex-specific checks often missed

• Test Credentials in Production
• Missing Input Validation
• Weak Auth Defaults
• Database Access Without Authorization