What Breaks When You Ship a Vibe-Coded App
AI tools can get an app to a working demo fast. Here's what tends to break after launch: leaked keys, open databases, payment gaps and store rejections.

AI coding tools are good at the first 80% of an app. You describe a screen, it appears, and by the weekend you have something you can show people. I use these tools in my own work, and I’d be slower without them.
The trouble starts when a demo meets real users. A demo only has to work once, on your phone, with your account and good Wi-Fi. A shipped app has to survive strangers, old devices, bad networks, app store reviewers and anyone curious enough to poke at it.
These are the problems I’d look for first in any app built mostly by prompting.
Secrets end up inside the app
This is the most serious one. An AI assistant asked to “connect to OpenAI” or “add Stripe” will often take the shortest route, which puts the secret key in the app’s code.
Anything shipped inside a mobile app can be extracted. Obfuscation slows people down and that’s all it does. A leaked AI key can run up a bill overnight. A leaked Stripe secret key is far worse.
The fix is a small backend or serverless function that holds the secret and makes the call for the app. Publishable keys can stay on the device if they are restricted properly. Secret keys can’t.
The database is open
Apps built on Firebase or Supabase are often generated with permissive rules, because strict rules make the demo fail. The app looks fine, while any signed-in user, or sometimes anyone at all, can read other people’s records.
Hiding a button in the app is not access control. The rule has to live on the server: Firestore security rules, Supabase row-level security, or checks in your API. I test this by signing in as one user and asking for another user’s data directly. If it comes back, the app isn’t ready.
Only the happy path exists
Generated code tends to assume everything goes right. Watch for:
- Screens that spin forever when the request fails
- Crashes when the API returns a null the code didn’t expect
- No handling for lost connections or expired sessions
- Forms that submit twice when someone taps twice
None of these appear in a demo. All of them appear in your reviews.
Payments work until they don’t
Taking a first payment is easy. The hard parts are renewals, failed cards, refunds, cancelled subscriptions, restoring a purchase on a new phone, and the webhook that arrives twice.
If the app grants access because the client said “payment succeeded”, someone will eventually get access without paying, and someone else will pay without getting access. The server should decide based on verified events from Stripe, RevenueCat or the store. I wrote more about that in my guides on RevenueCat subscriptions and Stripe Connect.
The same logic exists in five places
AI tools don’t always remember what they wrote an hour ago. So you get three date formatters, two API clients and a discount calculation pasted into four screens. Then you fix a bug in one place, and it stays broken in the other three.
This is what people usually mean when they say “every fix breaks something else”. The code is working by coincidence, and nobody can see how the pieces connect.
Nothing tells you when it fails
Most prototypes have no crash reporting and no useful logs. When a customer says “it doesn’t work”, you have nothing to look at. Adding crash reporting is cheap and should come before any other fix, because it tells you what to fix.
The stores push back
Apple and Google review what you submit, and vibe-coded apps run into familiar problems:
- Missing account deletion, which Apple requires if users can create an account
- Permissions requested with no clear reason
- Placeholder text, dead buttons or screens that go nowhere
- Sending personal data to an AI service without telling users and asking first
- A thin app that looks like many others, which can draw a Guideline 4.3 rejection
My release checklist covers what I check before submitting.
Is it worth fixing?
Often, yes. A prototype that proved people want the product has already done its most valuable job. The screens may be perfectly usable even when the plumbing underneath needs work.
Start with the things that can hurt you: secrets, data access and payments. Add crash reporting. Then fix what real users are hitting. Whether the rest deserves a cleanup or a rebuild depends on what you find, and I’ve laid out how I’d approach that in how I audit an AI-generated Flutter codebase. Crazybot Studio also has a guide for owners deciding whether to clean up or rebuild.
Takeaways
- Move secret keys off the device and rotate any that have shipped.
- Enforce data access on the server, then test it as a different user.
- Let the server decide who has paid.
- Add crash reporting before chasing individual bugs.
- Don’t rewrite anything until someone has read the code.
If you’ve built an app with AI and it’s starting to wobble, get in touch. Tell me what it does and what’s going wrong, and I’ll tell you honestly what I’d do first.
Comments
Questions, corrections or your own experience — leave a comment below (GitHub sign-in).