VIBE CODING

Vibe coding risks: is an app built with Lovable or Claude secure?

Vibe coding is building software by describing what you want to an AI tool, without writing the code yourself. With Lovable, Claude Code, Cursor or Bolt, something that works exists within days. The question that follows is whether you can trust it with real users and real data.

The short answer

Whether an app is secure does not depend on the tool it was built with. It depends on what was checked afterwards.

That is not a detail. In May 2025 a vulnerability was published (CVE-2025-48757) that let unauthorised people read and change data in apps generated with Lovable, because the access rules in the database fell short. Lovable disputes the report: according to the company, each customer is responsible for protecting the data in their own app.

Whoever is right, the lesson is the same: the tool delivers a working app, not the checks on it.

Read the CVE record

Wondering how your own app holds up? Send me the link on WhatsApp. Within one working day I'll tell you, free of charge, whether a review makes sense.

Where it goes wrong in practice

These are the points I look at first in an AI-built app.

  • Access to other people's data. The app checks who you are on screen but not on the server. Change a number in the address bar and you see someone else's records.
  • Keys in the code. Passwords and API keys sit in the source code or in the browser, where anyone can read them.
  • No tests. Nobody knows whether a new change breaks something that worked yesterday.
  • One file that keeps growing. AI tools happily keep writing in the same place. The result is a file in which neither a person nor the AI keeps the overview.
  • No separate test environment. Every change goes straight to real users, and the AI tool works against the real database.
  • No way back. There is a backup, but nobody has ever tried to restore it.
  • Outdated parts. The app leans on other people's packages that nobody updates.

Six things you can check yourself today

  1. Sign in as user A and try to reach user B's data through the address bar.
  2. Search your code for keys and passwords. If they are in the repository, replace them and take them out.
  3. If you use Supabase, check that access rules (row level security) are switched on for every table.
  4. Restore a backup into a test environment. Only then do you know it works.
  5. Ask yourself who gets a message when the app goes down at night.
  6. Check whether changes land somewhere else before they reach your real users.

If you get stuck on one of these, there is no reason to panic. It is a reason not to keep building as if everything is fine.

Can you keep vibe coding safely?

Yes. The speed of AI is not the problem, as long as the standard checks run automatically and a person looks at what matters.

In practice that means: each change in its own branch, automatic tests and scans on every pull request, a test environment with fictitious data, and AI tools that cannot reach the production database.

Instructions to the AI are not enough for this. What matters has to be enforced technically.

What a review cannot do

No review guarantees that software is free of errors or fully secure, and a baseline is not a certification such as ISO 27001. If you process sensitive data, an independent penetration test may be needed.

What you can achieve: knowing what is there, knowing which risks you accept, and having changes go live in a controlled way.

Unsure about your own app?

Send the link to your app. Within one working day you hear, free of charge, whether a baseline review makes sense, followed by a short call and a fixed price.

Send a WhatsAppContact form

A link, a screenshot or a voice message is enough. Usually a reply within one business day.

If an idea is unnecessarily complicated or expensive, I say so before it becomes code.