Vibe coding means describing an application in plain language, letting artificial intelligence write the code, and judging the result by what appears on screen rather than by the code itself. The method works remarkably well for shipping a prototype. On its own, it does not produce software a company can put online.
The gap between the two fits in a list. Secure sign-in, access rights, a structured database, backups, hosting, code review, automated tests, monitoring. This article covers what is left to build after the demo, what 2026 security surveys found in applications already online, and the four questions that decide the move to production.
What exactly is vibe coding?
Vibe coding is a way of producing software where the human describes the intent and AI writes all of the code. The term appeared in early 2025 from Andrej Karpathy, former director of AI at Tesla and an OpenAI cofounder. He described a practice where you accept the suggested code without reading it. The tool can be Lovable, Bolt, v0, Cursor or Claude Code.
The distinction matters, because two very different practices often share the same name. Addy Osmani, an engineer at Google, separates two practices. Vibe coding forgets the code exists. AI-assisted engineering keeps the engineer responsible for the architecture, reviewing every line produced. The first serves exploration. The second serves production.
Why the weekend demo is so convincing
A vibe-coded prototype reaches, in minutes, a level of polish that took several days two years ago. The interface is clean, the journey works, the business logic holds. The demo convinces, and rightly so, because everything visible really is finished.
I filmed that tipping point in a full tutorial where an interior rendering app starts from a single dictated sentence and runs locally after a few attempts. The limit shows up too, since the session ends on a list of what is still missing, the database and authentication first. The same demonstration shows up elsewhere, and I wrote about it in my OptimIA newsletter in November 2025, about a three-dimensional video game generated in a single flow.
That visible polish creates an illusion of progress. A prototype showing the right interface feels like an almost finished project, while the invisible half of the software has not started yet.
The nine levels below the waterline
What separates a prototype from production software fits in nine layers, all invisible from the screen. None of them requires technical genius. All of them require an explicit decision, and that is precisely what a prompt does not produce on its own.

- Secure sign-in. The app leans on the company's Microsoft or Google account rather than a password created for the occasion.
- Access rights. Each team sees only its own folders. This is what leaked the most applications in 2026.
- A structured database. One source of truth, no duplicates, with typed fields.
- Backups. One copy a day, and a restore that has been tested at least once.
- Controlled hosting. A company environment, with no API key written in plain text inside the code.
- Code review. By a human, or by an automated review tool wired to the repository.
- Automated tests. Nothing goes live without passing a set of checks.
- Monitoring. The team hears about an outage before users do.
- Handover. Someone other than the author must be able to change the app six months from now.
What 2026 surveys found inside vibe-coded applications
The first two levels on that list are the ones that give way first, and the measurements published in 2026 show it clearly. The problem rarely lies in the choice of tool. It lies in a database access rule nobody wrote.
The broadest survey comes from UpGuard. Looking at roughly 300,000 domains using the Supabase database, the team found 16,326 databases readable by anyone in September 2026. More than half of them held personal data, and over a thousand held credentials in plain text. The cause was an access protection that tables created by a coding agent do not switch on by default.

The clearest case involved a grading application showcased by the platform hosting it. A researcher found sixteen flaws there, six of them critical, and 18,697 user records reachable with no authentication at all, as The Register reported in February 2026. An inverted authentication function blocked signed-in users and let everyone else through.
The quality of generated code explains part of the phenomenon. Veracode, which has evaluated more than 150 large language models, measured in spring 2026 that 55% of generation requests produced secure code. The rest carried a known flaw. A model writes code that works long before it writes code that holds.
Four questions before putting a vibe-coded app online
Not every prototype belongs in production, and that is good news. A tool that never leaves the team that built it, touches no real data, and would cost nothing if it vanished can stay exactly what it is. Four questions settle it.

When no box is ticked, the prototype stays a prototype and nobody should spend three weeks hardening it. When a single one is ticked, the nine levels become mandatory before going live, and the project changes nature. Tandem treats that tipping point as a decision milestone, not a technical step.
Who builds the invisible levels, and how long it takes
The good news is that the list is finite and every line has an identifiable owner. The bad news is that none of those lines can be handed to a prompt. Here is the split Tandem applies on its internal tooling engagements.
| Level | Who owns it | Rough effort |
|---|---|---|
| Sign-in and access rights | IT team or partner | A few days |
| Database and backups | IT team or partner | A few days |
| Hosting and secrets | IT team | Under a week |
| Review and tests | Developer or review tool | Ongoing |
| Monitoring | IT team | One or two days |
This split assumes a rule written down somewhere, which is the job of clear AI governance. It says who may publish an application, and under what conditions. Without it, prototypes pile up and nobody knows which ones run in production. The same reasoning applies to AI agent security, where the risk comes from access granted without an inventory.
What Tandem puts into production, and what stays a prototype
Tandem works with more than 50 companies and has run around 100 engagements, a good share of which end in a custom internal tool. Our rule is stable. A vibe-coded prototype validates the idea, shows the screen to future users and settles an argument in a meeting. It never serves as the codebase for the version that goes live.
This site is one example. It was rebuilt with AI generation tools, and we told the full story in an article on rebuilding the site with AI. The code was still reviewed, versioned, tested and deployed through an automated pipeline. For teams who want to learn that discipline, the Claude Code guide for non-developers covers the basics.
Should companies ban vibe coding?
No, and banning it would mean losing the best scoping tool to appear in years. The right answer fits in two simple rules. Vibe coding is allowed everywhere to explore, mock up and convince. It stops at the waterline as soon as real data, an outside user or personal data enters the picture.
The companies doing best here are not the ones blocking tools. They are the ones that wrote down, on a single page, who can publish what and under which conditions. A prototype owned as a prototype causes no trouble. A prototype put online without anyone deciding causes trouble fast.



