Vibe coding security: a checklist before you launch your AI-built app

With vibe coding, you can build a working app in a weekend. You describe what you want, Lovable, Cursor, or Claude writes the code, and after a few rounds you have a customer portal or an internal tool. That is fast and useful. But the tool rarely asks on its own whether the app is safe enough for real users and real business data.

This checklist helps you answer that question before you launch. Each item has a check question. If you can answer it with confidence, move on. If you hesitate, you know where to look. You don’t need to be a developer to use the list. Some items are easier to check together with whoever set up the app technically.

A curved path of gray, blue, and cream stepping stones leading to an open door with blue light pouring out
Going live is a step. Check where you put your foot first.

What is vibe coding?

The term comes from Andrej Karpathy, one of the founders of OpenAI. In a post on X on February 2, 2025, he described a way of programming “where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.” You talk to the AI tool, accept its suggestions, and look at the result, not at the code. Collins named vibe coding its word of the year for 2025.

For a prototype, that is fine. It becomes a risk once the app goes live: people outside your team can log in, customer data goes in, or the app connects to a paid API. From then on, code nobody has read decides who gets to see what.

Why “it works” is not enough

An AI tool usually tests only the path where everything goes right. Veracode had more than 100 language models complete the same coding tasks: “45% of code samples failed security tests and introduced OWASP Top 10 security vulnerabilities into the code” (GenAI Code Security Report, July 2025). The code worked. It just wasn’t secure.

That is why you need a list that goes beyond “does the app do what I asked?” The ten items below start with your data, move on to access, and end with what happens after launch. To see what each flaw looks like in the code, also read the 7 most common security flaws in AI-generated code.

The checklist: ten questions before you go live

  1. 01

    Know what data your app stores

    Start with a short list: what data does the app store, where is it kept, and who needs it? Names, email addresses, invoices, and customer uploads are personal data. That means the GDPR applies, even to a small internal tool. Keep only what the app really needs. What you don’t store can’t leak.

    Check question: can you explain in two sentences which personal data your app stores, and why?

  2. 02

    No secrets in the code or the browser

    AI tools like to put an API key wherever the code works right away. That can be the source code, a .env file that ends up in the repository, or the JavaScript your app sends to the browser. Replace every exposed key with a new one at the provider, and keep secrets on the server. The Supabase documentation warns about this too: never use a secret key in the browser. GitHub secret scanning finds most keys in a repository.

    Check question: do you find a key when you search the repository, its history, and the JavaScript in your browser for words like key, secret, and token?

  3. 03

    The rules in your database are switched on

    Many vibe-coded apps talk to a database like Supabase or Firebase straight from the browser. That can be safe, but only if the database itself decides who can do what. In Supabase, this is called Row Level Security (RLS). The Supabase documentation is clear: “A table in an exposed schema without RLS is readable and writable by any role with a grant on it.” In Firebase, Security Rules do this job, and the open sample rules in the Firebase getting started guide are not meant for production.

    Check question: is RLS or a similar rule switched on for every table, and is it impossible to read data with only your app’s public key?

  4. 04

    Everyone sees only their own data

    Being logged in is not the same as having the right to the data. An app that only checks whether you are logged in will sometimes show another customer’s invoice as soon as you change a number in the link. The OWASP Top 10:2025 puts this flaw in first place, as Broken Access Control. The fastest way to test it is with two accounts.

    Check question: what does test account B see when it opens the links and API calls of test account A?

  5. 05

    Login and accounts are well protected

    Look at everything around the account, not just the login screen: the password reset, account creation, admin rights, and what happens after many failed attempts. Use the login of a well-known provider rather than a homemade one, and turn on two-factor authentication for admins. If you build the login yourself anyway, the details are in the OWASP Authentication Cheat Sheet.

    Check question: what happens after fifty wrong passwords, and can a new user make themselves an admin?

  6. 06

    Input and uploads are checked

    Anything a user types or uploads can be malicious: HTML in a name field, a script pretending to be a PDF, or a two-gigabyte file. Check input on the server, not only in the form in the browser. Limit uploads by type and size, and store them outside the place where your app runs code.

    Check question: what happens when you paste HTML into a text field, or upload an .html file instead of a photo?

  7. 07

    Every package exists and is up to date

    AI tools add packages quickly. Sometimes a package has a known vulnerability, and sometimes it doesn’t even exist. Let your package manager run a check, for example with npm audit or pip-audit, and look up who maintains every package you don’t recognize.

    Check question: does the check of your packages still report known vulnerabilities, and do you know every package on the list?

  8. 08

    Abuse won’t cost you a fortune

    If your app connects to a paid API, such as a language model, someone can call that API through your app thousands of times. You get the bill. OWASP calls this “Denial of Wallet” in its Top 10 for LLM Applications. Set a limit per user, set a budget cap with the provider, and route every call through your server. If your app has an AI feature, also watch out for prompt injection: text that makes the model do something it shouldn’t.

    Check question: what would it cost you if someone called your AI feature or your form 10,000 times tonight, and would you notice?

  9. 09

    You see errors, your visitors don’t

    A visitor should never see a stack trace, a database error, or an internal file path. You do want to know what went wrong. So log errors and suspicious events on the server, without passwords or full personal data in those logs, and make sure someone gets an alert. While you’re at it, check your security headers with a scan like Mozilla’s HTTP Observatory.

    Check question: what does a visitor see when something fails, and who gets an alert when something goes wrong?

  10. 10

    You can recover, and someone maintains the app

    Going live is not the finish line. Set up automatic backups of your database in a separate location, and try once to restore from them. Agree on who updates the app when a flaw in a package becomes known. And who decides when there is a data breach? Under the GDPR, once you discover a breach that puts people at risk, you have 72 hours to report it to the supervisory authority. In Belgium, that is the GBA/APD. In the Netherlands, it is the Autoriteit Persoonsgegevens.

    Check question: if your database were empty tomorrow, how would you restore it, and who updates the app when a flaw becomes known?

What can you check yourself, and where does a second pair of eyes help?

You can check off part of the list yourself in an afternoon. For other items, someone needs to read the code and know how the app is used. This table helps you decide.

ItemCan you check it yourself?Where a second pair of eyes helps

Data and GDPR

Yes, with a short list per table

With sensitive data, such as health or HR records

Secrets

Yes, with a secret scanner

If a key was ever in the repository history

Database rules

Partly: check that RLS or Security Rules are on

To verify the rules are right for every role

Own data only

Yes, with two test accounts

With roles, teams, or several companies in one app

Login and accounts

Partly

With a homemade login

Input and uploads

Partly

With customer uploads or documents an AI feature reads

Packages

Yes, with npm audit or pip-audit

With unfamiliar or rarely used packages

Costs and abuse

Yes: set limits and a budget cap

With AI features that take actions on their own

Errors and logs

Yes, with a header scan and a triggered error

To decide what you do and don’t log

Recovery and maintenance

Yes, by restoring from a backup once

If nobody really knows the code

Do many of your items end up in the right-hand column? That is not a bad sign. It means your app does something real, with roles, customer data, or an AI feature. That is exactly where an AI tool misses things most often, because it doesn’t know your business rules.

What do you do with an item you can’t check off?

Don’t open the app to everyone yet. These in-between steps help:

  • Launch to a small, known group first, with test data instead of real customer data.

  • Ask your AI tool about one item at a time, for example: which tables can be read without logging in? Then check the answer yourself.

  • Use the web application security guidelines from the Dutch NCSC as a shared framework with whoever built the app.

  • Put the list next to the OWASP Top 10, so your developer can look up what each item means.

  • Have the code reviewed before real customer data or payments go in.

No checklist makes an app 100% secure. A checklist does remove the known gaps before you launch. To learn when a full review makes sense, read why an app built with AI needs a security audit. Voltti does security audits for businesses in Belgium and the Netherlands. Want someone to examine your app thoroughly? Take a look at the security audit by Voltti. AI models do the broad search work there, and a person with 20+ years of experience in web development reviews every finding. You can read how that works step by step in how Voltti runs a security audit with AI.

Portrait of Seppe Gadeyne

Seppe Gadeyne is a developer and the owner of Voltti. He sets up AI agents for companies in Belgium and the Netherlands.

Updated on

Frequently asked questions

01What is vibe coding?

Vibe coding means building software by telling an AI tool in plain language what you want, and accepting the suggested code without reading it line by line. Andrej Karpathy gave the concept its name in February 2025. With tools like Lovable, Cursor, and Claude, you can build a working app this way in a short time.

02Is vibe coding safe?

For a prototype with test data, the risk is small. If the app goes live with real users or business data, someone needs to check what the code does. AI tools write code that works, but they rarely check on their own who can see which data or where keys end up.

03Do I need to know how to code to use this checklist?

No. Several items, such as the two-account test, a triggered error message, or a header scan, can be done without reading the code. For items like the rules in your database and keys in the repository history, it’s best to work with whoever set up the app technically.

04Is my app secure if I can check off every item?

Then you have covered the best-known risks, but it is not a guarantee. Flaws in your app’s logic don’t always show up in a checklist. If the app handles customer data, payments, or sensitive documents, have the code reviewed as well.

05When should I have a vibe-coded app reviewed?

As soon as people outside your team can open the app, or real business data goes in. A review also makes sense after a major change by an AI tool, such as a new payment integration or a new permission structure.

Want your AI-built app reviewed before launch?

Discuss your pilot