launch■blockers

> the 15 ways AI-built apps leak their users' data, checked before you launch

Built your app with Lovable, Bolt, v0, Replit, Cursor or Claude Code? These are the mistakes behind the leaks in the news: open databases, keys in the browser, Firebase left in test mode. Every pattern links to a dated incident or the platform's own rule (11 from dated incidents or studies, 4 from the platform's own rules), comes with a fix prompt you paste into your builder, and has a check that runs before you deploy.

$ npx launch-blockers check
═ before the deploy ═
$ npx launch-blockers check
launch-blockers · BLOCKED

2 things to fix before you launch:

1. The secret VITE_OPENAI_API_KEY is named so it ships to the browser  [critical]
   src/ai.ts:1
   Fix prompt: Rename it without the prefix, move every use of it into backend code…

2. The table "profiles" is created without Row Level Security  [high]
   supabase/migrations/20260101_init.sql:1
   Fix prompt: Write a new migration that enables it and adds policies…

Not checked (CLEAR never covers these):
  - Your live database: npx greenlitlab <your-app-url> --mine

$ # fix, then check again
launch-blockers · CLEAR
The sample project in the repo, before and after its two fixes. Everything is read on your computer; nothing is uploaded.

═ why this exists ═

None of these needed a hacker. Each was one setting.

═[1]═ the 15 patterns

Each one has a source anyone can read. Numbers are the source's own, and caveats stay next to them in the playbook.

#patternseverityfound byon the record
1Database tables with no Row Level SecuritycriticalscanCVE-2025-48757: Lovable apps without working RLS (2025-05)
2Row Level Security is on, but the rule checks nothinghighscanCVE-2025-48757: Lovable apps without working RLS (2025-05)
3Firebase rules that let anyone in ("if true" or test mode)criticalscanAI chat app exposes chat histories (2026-01)
4File storage anyone can read (ID photos, documents, uploads)criticalscanTea app: verification selfies and IDs exposed (2025-07)
5A paid API key (OpenAI, Gemini, Stripe...) shipped in the browser codecriticalscanAbout 3,000 websites exposing ChatGPT keys (2026-02)
6The Supabase service_role (secret) key in the browsercriticalscanRule: Supabase docs: API keys
7The .env file committed to githighscan5,000+ GitHub repositories with OpenAI keys committed (2026-02)
8A backend function anyone on the internet can callhighscanRule: Supabase docs: function configuration (verify_jwt)
9A payment webhook that believes anyonemediumscanRule: Stripe docs: verify webhook signatures
10An access check that's backwards or only in the browsercriticalagent reviewLovable-showcased app with an inverted access check (2026-02)
11Change the number in the address, see someone else's datacriticalagent reviewRule: OWASP API Security Top 10: API1 Broken Object Level Authorization
12A private app that still lets anyone sign uphighagent reviewBase44: anyone could join private apps (2025-07)
13An internal tool published to the open internethighagent reviewRedAccess: thousands of public vibe-coded apps with company data (2026-05)
14Your coding agent can change the production databasemediumscanReplit agent deletes a production database (2025-07)
15The live database isn't what your files sayhighlive checkCrissy Field scan of 159,041 Lovable projects (2026-11)

found by: scan = launch-blockers check finds it in your files · agent review = the skill walks your coding agent through it · live check = only visible on the running app.

═[2]═ the deploy guard

One command adds a skill and a guard to Claude Code. From then on, a deploy waits for the check.

$ npx launch-blockers install
BLOCKED
The deploy is stopped. Your agent gets the reasons and the fix prompts, so it can fix them and try again.
NOT CHECKED
The check couldn't run, so you're asked. It never passes silently.
CLEAR
Nothing changes: your usual approval still applies, with a note of what wasn't checked.

It watches vercel, netlify deploy, supabase db push, supabase functions deploy, firebase deploy, wrangler deploy, fly deploy, railway up and npm run deploy, only when they're the command being run. A commit message that mentions a deploy is left alone.

Checked a finding yourself and it's fine, say Row Level Security switched on in the dashboard? Add its id and your reason to .launch-blockers-ignore. A line without a reason doesn't count, and ignored findings are always listed.

═[3]═ fix prompts

Most people building with AI fix things by asking. So every pattern comes with the exact request to paste into Lovable, Bolt, Cursor or Claude Code.

# pattern 3 · Firebase rules that let anyone in
Rewrite my Firebase security rules so no rule uses if true or the
test-mode date check. A signed-in user should reach only their own
documents (request.auth != null && request.auth.uid == the owner
field or path segment). Keep anything truly public read-only and tell
me which it is. Show me the new rules and how to test them in the
Firebase emulator before deploying.

Not a developer? Open the playbook and work down the list. Ten minutes, no terminal.

═[4]═ what it can't see

═[5]═ faq

Does my code get uploaded?

No. Everything runs on your computer. Names in .env files are read, never values.

Does it work with Lovable or Bolt, where I don't use a terminal?

Yes, through the playbook: each pattern's fix prompt is written to paste into your builder. To check the live app from outside, use Greenlit, our free grade for AI-built apps.

Which coding agents does the guard work with?

The deploy guard is for Claude Code today. The skill file works in any agent that reads skills, and launch-blockers check works anywhere Node 20 runs, including CI (exit code 1 when blocked).

Can I check someone else's app?

No. Check only what you own. The live check reads only what the app's own public key can reach, and still, only point it at your own app.

How do I add a pattern?

Open an issue with a public link: an incident write-up, a study or the platform's own documentation. Patterns go in only with a source anyone can read.

Who makes this?

Priberry. The scanning is Greenlit's project scan. Apache-2.0.

═[6]═ get started

Free and open source. Needs Node 20 or newer, nothing else.

$ npx launch-blockers check
BLOCKED, CLEAR or NOT CHECKED for the project in this folder. Add --json for CI.
$ npx launch-blockers install
The skill and the deploy guard for Claude Code, in this project. Add --global for every project.
$ npx launch-blockers list
Every pattern, with the incident behind it.
$ npx greenlitlab yourapp.com --mine
The live check, for your own app.

Version 0.1.0. Tell us what breaks: hello@priberry.com · github issues