Insights / October 20, 2026

Vibe coding security: the risks we find in AI-built apps

AI app builders are very good at making things work. They’re much less reliable at making sure things only work for the right people. That gap is where nearly every security problem in a vibe-coded app comes from.

This isn’t theoretical. In 2025 a researcher scanned 1,645 apps built with Lovable and found 170 of them letting anyone read or change their databases, recorded as CVE-2025-48757. Lovable disputes the CVE, arguing that each customer is responsible for their own app’s data. That’s technically fair, and it’s also the point: the tool won’t protect your users for you.

Here are the risks we find most often when founders bring us AI-built apps, roughly in order of how much damage they do.

1. The database has no real rules

Most AI-built apps keep their data in Supabase. Supabase protects data with row-level security: rules on each table saying who can read and change which rows. Your app’s frontend talks to the database directly, so these rules are the only thing standing between one user and everyone else’s data.

When the rules are missing, or say true, anyone who opens their browser’s developer tools can read the whole table. That’s what the CVE above was about.

How to check: in Supabase, confirm row-level security is on for every table, and that each rule matches who should really have access. Our guide to Supabase row-level security walks through it.

2. Security checks that only happen in the browser

AI tools often hide what a user shouldn’t do instead of preventing it: the admin button doesn’t show, the paid feature is greyed out. But the browser is under the user’s control. If the server doesn’t check too, anyone can call the action directly.

How to check: for every admin or paid action, ask where the check happens. If the answer is “the button is hidden,” it isn’t checked.

3. Secrets in the frontend

Apps call paid services like OpenAI, Stripe and email providers, and those calls need secret keys. AI tools sometimes put the keys in code that ships to the browser, where anyone can copy them. The worst one is Supabase’s service_role key, which skips every database rule.

How to check: search your frontend code for service_role, sk_ and secret. Anything that looks like a key belongs in a server function, never in the browser.

4. Server functions anyone can call

Edge functions, API routes and Next.js server actions run on a server, so they feel safe. But they’re public URLs. If one doesn’t check who’s calling it, anyone can, as often as they like. When the function calls a paid AI model, that means a stranger can run up your bill.

How to check: every server function should require a signed-in user unless it’s genuinely public, and anything that costs money per call needs a rate limit.

5. Payments that can be faked

Stripe tells your app about payments through webhooks. If the app doesn’t verify that a webhook really came from Stripe, anyone can send a fake “payment succeeded” message and unlock a paid plan. AI-built checkouts often skip this step because the happy path works without it. More in adding Stripe to an AI-built app.

How to check: find your webhook handler and confirm it verifies Stripe’s signature before trusting anything in the message.

6. Files anyone can open

Uploaded files often land in a public storage bucket, or get shared through links that never expire. That’s fine for a logo and a serious problem for contracts, IDs or private photos.

How to check: look at each storage bucket’s access setting, and make sure private files use signed links that expire.

7. Packages nobody chose

AI tools add libraries freely. Some are outdated, some are unmaintained, and researchers have shown that AI models sometimes suggest packages that don’t exist, which attackers can then register under that name.

How to check: run your package manager’s audit command, and look up any package you don’t recognize before it reaches production.

8. AI features that can be talked into things

If your app has its own AI features, like a chatbot that reads your data or an agent that takes actions, it can be manipulated by text hidden in what it reads. This is called prompt injection, and there’s no complete fix. The defense is limiting what the AI is allowed to see and do.

How to check: list what your AI features can access and change. Each one should have the least access it needs, and anything destructive should need a human to confirm.

Why AI tools create these problems

None of this is because the tools are bad. They optimize for what you can see: does the button work, does the page load. Security is mostly about what you can’t see, the requests nobody is supposed to make. A demo never makes them, so the code that blocks them never gets asked for.

A quick self-check

If you only have an hour:

  1. Turn on row-level security for every Supabase table and read every rule.
  2. Search your frontend code for secret keys.
  3. Sign up as a new, ordinary user and try to reach admin pages and other users’ records.
  4. Confirm your Stripe webhook verifies signatures.
  5. Make sure any function that calls a paid API requires sign-in and has a limit.

If that hour turned up more than you expected, a vibe coding rescue starts with a full security pass, so you know what to fix before real users find it.

LET’S TALK.

Book a free 30-minute call with Virgil

Pick a time that suits you. You’ll get a straight answer on what you need, what it should cost and the fastest path to launch.

Loading available times…

Open the calendar in a new tab

Send us your project

Virgil replies personally within one business day.

Book a free strategy call