Vibe coding: the app works. Can it go live?
An employee describes in plain language what they want, and an hour later there is a working app: a customer portal, a time tracker, a quoting tool. That is vibe coding, and it really is that fast. What the screen does not show is who else can reach the data, what happens when something goes wrong, and who maintains the app six months from now. That is the difference between a prototype and something customers rely on.
What vibe coding is
The term comes from Andrej Karpathy, former AI director at Tesla and a co-founder of OpenAI. In February 2025 Karpathy described a way of building where you let the AI write the code, judge the result on whether it works, and forget that the code even exists. It was meant for weekend projects.
More than a year and a half later it is a category of tools. Lovable, Bolt and Replit build a complete web app with a database and login from a description. Cursor and Claude Code do the same for people who do work in code. The result looks professional and does what was asked. So people who could not build software before now do: the operations manager, the controller, the business owner.
What it is good for
Vibe coding is not a hype to dismiss. Testing an idea no longer takes a proposal and months of work. Within a day you see whether a customer wants to use something, whether an internal process really gets simpler, or whether the idea sounded better on paper. A clickable prototype says more than a twenty-page specification, including to a developer who builds it later.
For a demo, an internal tool without sensitive data, or a first version to gather feedback, it is often exactly enough. The risk starts when the prototype quietly becomes production: real customers, real data, and nobody who decided it was ready.
What went wrong in practice
Two examples from 2025 show where it pinches. Neither is a beginner's mistake; they are failures the app itself does not make visible.
In the spring a security researcher examined apps built with Lovable. In 170 of the more than 1,600 apps examined, database tables could be read or changed without logging in, exposing names, addresses and payment details among other things. The app had a login screen, but the database behind it did not check who was asking. The flaw received an official CVE number.
In July the founder of SaaStr spent twelve days publicly building an app with Replit's AI agent. On day nine the agent deleted the production database, while a freeze on changes was in place. Test and production used the same database. Replit has since separated the two by default.
The pattern is the same in both cases: the AI builds what it is asked, and the questions nobody asks go unanswered. People who are not developers often do not know those questions exist.
Five questions before an app goes live
If a vibe-coded app is going to serve real customers or hold real data, these questions need answers. If nobody can answer them, that is the answer.
- Who can see what? Not on the screen, but in the database and the API behind it. Try to request another user's data. If that works, the login is decoration.
- Are there secrets in the browser? API keys, database keys with full rights or passwords belong on the server. Anything in the browser can be read by anyone.
- Are test and production separate? An AI agent that is allowed to experiment should not be able to reach the real data. And there should be a backup you have actually restored at some point.
- Where is the data, and with whom? The build tool, the database and the hosting are each a supplier. With personal data you need data processing agreements, and sometimes a DPIA.
- Who owns it? Someone has to apply updates, follow up on flaws reported in components it uses, and know how the app works when the person who built it leaves.
It is also wise to add the app to your AI register, or at least to your list of software in use. A tool an employee built themselves that processes customer data is exactly the shadow IT that comes up first after an incident.
What Axon Factory adds
We build with AI agents every day ourselves. The difference from vibe coding is not the tools but what surrounds them: a second agent or person who checks the work with the assignment to find mistakes, tests on what a user sees, recorded decisions, and a person who decides about production. We bring that along when you come to us with a vibe-coded app.
- An assessment of what is there. A consultant checks the app against the five questions above: database access, keys, separation of environments, suppliers and maintenance. You get a report with findings and what needs to happen before the app can go live.
- Making it production-ready, or rebuilding it. Sometimes the prototype is a good base and we fix the findings. Sometimes the honest advice is to use the prototype as the specification and set up the app again. Then we say so.
- The agreements around it. Who may build what with which tools, which data may go in, and when someone should take a look. With AxonVendor you assess the suppliers behind the app, with AxonDPIA you check whether a DPIA is needed.
The goal is not for employees to stop building. The goal is that a good idea does not fail on a flaw that could have been found in an afternoon.
Sources: Andrej Karpathy's post on X (2 February 2025), CVE-2025-48757 on insufficient Row Level Security in apps generated by Lovable, and reporting on the Replit incident of July 2025, including the response from Replit CEO Amjad Masad. The Lovable figures come from the discoverer's research as reported by security firms.