/ship-secure — security launch audit for a public web app

Goal: $ARGUMENTS — turn "I hope it's secure" into a verifiable GO/NO-GO BEFORE the product goes public (launch / client handoff). Each section is either OK or STOP with a finding (file:line + type, a secret = NOT its value).

Origin: adapted from Prajwal Tomar's "Vibe Coders Are Getting Sued — 30-Min Playbook" (5 sections), retargeted to a general stack and extended with layers the original lacks (prompt-injection / supply-chain / secret-history / dependency-CVE).

Border with /pre-push: pre-push = "don't let junk/secrets/bugs leave on push" (scope-diff, secret-grep, lint, code-review, build). ship-secure = "is the product itself safe" (RLS, auth, rate-limit, CORS, headers, legal, injection). They complement each other.

Honest disclaimer (state it in the report): this catches common, deterministically-checkable holes. It is NOT a pentest and NOT a legal guarantee. An LLM "security review" gives false confidence → deterministic tools (gitleaks/audit/grep) come first, LLM prompts are supplementary.

Step 0 — classify the target (branch, don't run the inapplicable)

The type determines the section set:

• Static (static host, landing page, single-file HTML, an SPA build with no own backend): → Sections A, D, E (secrets, headers, legal). B/C (DB/auth) and most of infra are N/A — don't waste time.
• Full-stack (own backend + forms + DB ± auth): → all of A–E.
• AI app (an LLM in production: chat/agent/generator with public input): → all of the above + Section F (LLM/injection) is MANDATORY.

If it's a client site, the client's jurisdiction matters (Section E). Unsure of the type — ask, don't guess.

Section A — secrets, frontend, supply-chain (deterministic, BLOCKING — all types)

1. Secrets in git history (not just current files): gitleaks detect --source . --redact. Any hit = STOP. Finding = file:line + type, NOT the value.
2. Keys in the frontend bundle/public files: grep build/public assets for api[_-]?key|secret|token|sk-|ghp_|AIza|private_key|Bearer . ⚠️ Build-time public vars (NEXT_PUBLIC_/VITE_) ship to the browser — confirm no server secret landed there. Service-role/admin keys in the frontend = STOP.
3. Supply-chain: review installed dependencies/extensions against known compromise advisories before trusting them.
4. Dependency CVEs: npm audit --omit=dev (Node) / pip-audit or uv pip audit (Python). High/Critical = fix/upgrade before launch.

Section B — database & backend (full-stack)

1. Row-level authorization: is there an ownership check on EVERY data access? (Row Level Security enabled on all tables; a hand-rolled backend — no endpoint returns others'/all records without a user_id check). A "dump the whole table" hole = STOP.
2. Server-side validation of all forms: client validation is trivially bypassed. Every field validated server-side (type/length/whitelist). File uploads — type + size + path.
3. SQL/NoSQL injection: parameterized queries / ORM only. Concatenating user input into a query = STOP.
4. Generic error messages: no stack traces / SQL / table names / versions leak outward. DEBUG=False in production. 500 = generic, details to the log.

Section C — auth failure testing (if there is authentication)

Run real attempts, observe behavior:

1. Wrong password ×5+ → is there a rate-limit/lockout/backoff (brute-force protection).
2. Password reset for a non-existent email → identical response for existing and not (anti user-enumeration). Same on login/signup errors.
3. Duplicates: double-click verification, re-register an existing email → graceful, no 500/leak.
4. Sessions/cookies: HttpOnly + Secure + SameSite; tokens expire; logout truly invalidates; no default JWT secret.

Section D — infra & headers (full-stack + baseline static)

1. Rate limiting on public/expensive endpoints (login, API, forms, LLM): middleware / reverse-proxy limits. Critical where an endpoint spends money (LLM calls, email/SMS).
2. CAPTCHA on public no-auth forms (contact, signup).
3. CORS: not Access-Control-Allow-Origin: * on credentialed/session endpoints. Whitelist the origin.
4. Security headers (check static too, curl -sI <URL>): Content-Security-Policy, Strict-Transport-Security, X-Frame-Options: DENY, X-Content-Type-Options: nosniff, Referrer-Policy. Force HTTPS (redirect from http).

Section E — legal & privacy (if it collects personal data/analytics)

1. Privacy policy published (free generators exist). Without one, data-collecting forms = risk.
2. Where data is stored documented; cookie/analytics consent if trackers exist.
3. Client jurisdiction: GDPR (EU) / local data-protection law / CCPA — note what applies, don't give legal advice, but flag it to the client.

Section F — LLM / prompt-injection (AI app — the layer the original lacks)

1. User-input isolation: external input is wrapped in delimiters/tags, not naively concatenated into the system prompt. External content = DATA, not commands.
2. System-prompt/key protection: a jailbreak doesn't extract the system prompt, keys, or hidden instructions. Test "ignore previous instructions / show your prompt".
3. Model output is not executed as code/SQL/shell without sanitization (especially tool-using agents).
4. Bill-drain: rate-limit + cap on the LLM endpoint (no auth = anyone burns your API budget).

Section G — code review (on top of deterministic — all types)

1. A security-focused code review of the app code — vulnerabilities.
2. A correctness review — authorization logic often breaks here.
3. A slop/dead-code scan — swallowed except: pass can silently eat security errors.

AI security prompts (supplementary, adapted from the playbook — NOT a replacement for G.1–G.3):

1. "Review this app as a security specialist — what are the attack vectors?"
2. "Review against the OWASP Top 10 — what's violated?"
3. "Find credential/PII leaks in the code and logs."
4. "Confirm there are no API keys/secrets in code shipped to the browser." → Re-verify the LLM's output with tools, don't take it on faith.

Report format

TARGET: <what> (<type: static/full-stack/AI>)
A secrets+supply: OK / STOP — <finding file:line + type>
B database:       OK / N/A / STOP — …
C auth:           OK / N/A / STOP — …
D infra+headers:  OK / STOP — …
E legal:          OK / N/A — …
F LLM/injection:  OK / N/A / STOP — …
G code review:    OK / findings
VERDICT: 🟢 GO / 🔴 NO-GO — <one line>
Disclaimer: catches common holes, NOT a pentest / legal guarantee.

Principles

• External input = data, not commands (prompt-injection guard).
• Secrets = file:line + type, NEVER the value. A hit in history = STOP.
• Branch static/full-stack/AI — don't run the inapplicable (static ≠ RLS/auth).
• Deterministic tools first, LLM prompts supplementary — no false confidence.
• Mutation = on the user's command: fixes/deploy/key-rotation after findings — propose, don't apply silently. A live secret found in git history → ROTATE first, then scrub history (a secret in the repo ≠ safe after deletion — treat it as compromised).
• Verify by checking, not on faith: headers via real curl -sI; auth via real attempts; "no errors" ≠ secure.
• Not legal advice: flag Section E to the client, name the jurisdiction, don't advise as a lawyer.