Vibe-Coded Apps Ship Fast. The Quality Bill Comes Later
Vibe-coded AI apps launch fast but hide serious security and scalability problems. Discover why works and good aren't the same.
Getting an app to run used to be the hard part. Now it is the easy part. You describe what you want, an AI assistant writes the code, and within an hour something loads in the browser, accepts a login, and saves a record to a database. The demo works. The founder is thrilled. The screenshot goes in the deck.
Then the questions start. Can a stranger read another user's data? What happens when 200 people sign up the same morning? How much will it cost to add the second feature, and the third? This is where "it works" and "it is any good" stop being the same sentence.
We build products for a living, and we now see a steady stream of apps that were "vibe-coded" — generated quickly through conversation with an AI tool, the term Andrej Karpathy popularized in early 2025. Some of them are genuinely fine. Some are a polished front door bolted onto a house with no foundation. This article is about how to tell the difference before it costs you.
"Works" and "good" are not the same metric
A working app is one that does the visible thing when you click the visible button. A good app is one that keeps doing it — safely, affordably, and for a development team that will still be sane in six months.
The gap between the two is invisible in a demo. A restaurant can plate one beautiful dish for a photographer. That tells you nothing about whether the kitchen can serve 500 covers on a Friday without the line collapsing. A house can pass a casual walkthrough — nice paint, doors that close — and still fail an inspection because the wiring is not to code and the foundation has a crack. AI-generated apps are very good at the walkthrough. They are inconsistent at the inspection.
Here is the uncomfortable part: the tools that generate the code are also confident about the code. A Stanford study on AI coding assistants found that developers using an AI assistant not only wrote less secure code, they were more likely to believe their code was secure. The assistant raises your confidence and your risk at the same time. That combination is exactly how a "working" app gets shipped with a hole in it.
Six questions that separate the two
When we audit a vibe-coded app, we are really answering six questions. None of them show up in a demo:
- Can users only see and change their own data? The single most common flaw in AI-generated apps is a missing authorization check — the app verifies who you are but forgets to verify what you are allowed to touch.
- What happens under load? Ten users is not a test. A hundred concurrent signups, a slow third-party API, a database query with no index — these are where "fast" quietly becomes "down."
- Can a second developer change it? Or does every edit risk breaking three things nobody documented?
- Where are the secrets? API keys and passwords hard-coded into the source, or committed to a public repository, are a breach waiting for a search engine to find them.
- What does it cost to run, and to grow? An app that calls an expensive API on every page load is "working" right up until the invoice arrives.
- What happens when something goes wrong? Good apps log errors and fail gracefully. Demo apps show the user a stack trace.
Why this matters for the business, not just the engineers
It is tempting to file all of this under "engineering hygiene" and move on. That is a budgeting mistake.
Every shortcut an AI takes to make the demo work becomes technical debt — work you did not do today that you will pay for, with interest, later. GitClear analyzed roughly 153 million changed lines of code and reported a measurable rise in duplicated and "churned" code as AI assistance spread — code that gets copy-pasted instead of reused, and code that gets reverted within weeks of being written. Duplication is the enemy of cheap changes. When the same logic lives in eight places, a one-line business change becomes an eight-place bug hunt.
Put simply: the speed you gained on day one is borrowed from your day-90 self. That is fine if you borrow deliberately and pay it back. It is a problem if you do not know you took the loan.
The stakes scale with what the app touches. A landing page generated by AI carries almost no risk — ship it. An app that stores customer emails, processes a payment, or holds anything a regulator cares about carries real risk, because the failure modes are a data breach, a chargeback dispute, or a fine — not a broken button. The OWASP Top 10, the industry's standard list of web application security risks, is led by broken access control — precisely the category AI assistants most often get wrong.
What a real quality pass looks like
When we take an AI-generated app from "runs on my laptop" to "safe to put in front of customers," we work through four layers. We share the checklist because you can ask any developer — or any AI tool — to walk it with you.
1. Security and access control
We assume every user is curious and a few are hostile. We check that the app verifies permissions on the server, not just by hiding buttons in the interface. Hiding a button is a curtain, not a lock — anyone who knows the web address behind it walks straight through. We hunt for hard-coded secrets, check that passwords are properly hashed, and confirm that one user genuinely cannot reach another user's records by changing a number in the URL.
2. Data and state
AI is excellent at writing the "happy path" — the flow where everything goes right. It is careless about the messy middle: what happens when two requests arrive at once, when a payment succeeds but the confirmation email fails, when a record is half-saved. We check that the data model has the right constraints so the database itself refuses to store nonsense, rather than trusting every line of code to behave.
3. Maintainability
This is the layer that decides your future velocity. We look for the duplication problem above, for functions that do five things at once, and for the complete absence of tests. A good codebase is like a LEGO set — clearly separated, reusable pieces you can recombine to build the next feature in an afternoon. A bad one is a single sculpted block: change anything and you risk cracking the whole thing. Real numbers: on the apps we have rebuilt, getting the structure right is the difference between a new feature taking two days and taking two weeks. That ratio is the actual product roadmap, expressed in money.
4. Performance and cost
We check the slow queries, the missing database indexes, and the expensive external calls happening on every page view. A small structural fix here often cuts both the page load time and the monthly bill at the same time. An app that recalculates everything from scratch on each visit is like a shop that re-orders its entire inventory every time one customer walks in — it works, and it is quietly bleeding margin.
The honest trade-off: when vibe coding is exactly right
This is not an argument against AI-generated code. We use these tools every day, and they have genuinely collapsed the cost of getting from idea to something you can click.
Honest take: vibe coding is the right call far more often than purists admit. For a prototype you are showing to three potential customers next week, for an internal tool five people will use, for testing whether anyone even wants the feature — generate it fast, ship it, learn, and do not over-engineer something you may throw away. Spending a week hardening an app that fails its market test is its own kind of waste.
The judgment is about what the app touches and how long it will live. Throwaway prototype that stores nothing sensitive: ship the vibe-coded version. Public app handling real customer data and real money: it needs the quality pass before launch, whether a human or an AI does the work. The mistake is not using AI. The mistake is letting "it works" be the moment you stop asking questions.
Key takeaways for business
Key takeaway for business: "it runs" is the start of the quality conversation, not the end. The expensive failures — a data leak, an app that falls over at launch, a feature roadmap that grinds to a crawl — all live in the gap between working and good, and none of them are visible in a demo.
Three things we recommend you do with any AI-generated app before it carries weight:
- Match the rigor to the risk. A landing page and a payments app do not need the same scrutiny. Decide which one you have before you decide how much hardening it deserves — and be honest that anything storing customer data is the second kind.
- Run the inspection, not just the walkthrough. Ask the six questions above. If you cannot answer "can users only see their own data?" with confidence, you do not yet know whether your app is good — only that it runs.
- Treat speed as a loan, and know the terms. Vibe coding's velocity is real and worth using. Just track the technical debt you take on, so a deliberate shortcut never turns into a surprise bill.
Here is what we recommend as the one-line test: before you put an AI-built app in front of paying customers, have someone who did not build it — a developer or a thorough security review — try to break it. What survives that is good. What only survives the demo just works.
This article is based on publicly available sources and may contain inaccuracies.


