A weekend is enough. A few hours of conversation with an AI tool, and you have an app with sign-up, a dashboard and a simulated checkout that impresses everyone in Monday's meeting. That is a real achievement, and we have no quarrel with it. The trouble starts when the same version, unchanged, meets real users, real data and real attackers.
This guide is for CEOs, founders and product leaders who have an AI-built prototype in hand, or whose teams write much of their code with AI, and who now need to decide how to turn it into a product the business can rely on. We explain what vibe coding actually is, why its output works in a demo but breaks in production, what the latest evidence shows, and what the path from prototype to production system looks like.
Summary for managers
- Vibe coding means building software by describing what you want to an AI, without reading or understanding the code it produces. It is excellent for prototyping and testing ideas; it is not enough for a system that holds users' data and money.
- Vibe coding's standard is "it works". Production's standard is that it works securely, stays up, can be monitored and can be maintained, and most of that is invisible in a demo.
- According to Veracode's latest data, AI-generated code still introduces vulnerabilities in more than a third of security tests.
- The answer is not to abandon AI. Generated code has to pass the same engineering gate as any other code: review, tests, security scanning, separate environments, monitoring and a clear owner.
What is vibe coding?
The term vibe coding was coined in February 2025 by Andrej Karpathy, an OpenAI co-founder and former head of AI at Tesla. He described a way of working in which the programmer tells an AI what they want in plain language, accepts its changes without reading them, and pastes any error message straight back into the tool, to the point where, in his words, you forget the code even exists. The term spread so fast that Collins Dictionary named it its Word of the Year for 2025.
One part of that definition usually gets dropped: in the same post, Karpathy said the approach was not too bad for "throwaway weekend projects". The person who coined the term pitched it for fast prototyping, not for systems a business depends on.
Martin Fowler draws the same line in his note on vibe coding: vibe coding is building software by prompting an LLM without looking at the code it generates. That is different from AI-assisted programming, where an engineer uses AI to move faster but reads, understands and takes responsibility for every change.
| Vibe coding | AI-assisted engineering | |
|---|---|---|
| Who reads the code | Nobody, or only the tool | The engineer responsible for the change |
| What counts as correct | "It works on screen" | Automated tests, review and behavior in production |
| Security and access | Whatever the tool's defaults are | Designed and reviewed |
| Good for | Prototypes, idea demos, personal tools | A product that holds users' data, money or trust |
The difference has nothing to do with how much AI is used. A team that has AI write most of its code can work in a fully engineered way, and someone who vibe-coded a single simple form can leave every user's data open to the public. The real question is whether anyone has understood the code and taken responsibility for it.
Why the demo works and production doesn't
An AI tool optimizes for what you can see and give feedback on: does the page open, does the button work, does the data show up? Anything you cannot see has no place in that feedback loop.
The problem is that most production requirements are exactly that kind of thing. A demo has no user who must not see another user's data. It has no thousand concurrent requests, no external service going down, no attacker calling the API directly, and no developer who has to add a feature six months from now. A prototype built only to the standard of "it works" is empty in precisely those invisible places.
The evidence from the past eighteen months
This concern is not theoretical. A few public data points and incidents show the pattern clearly:
- Security of generated code. Since July 2025, Veracode has been testing code generated by more than 100 language models on 80 coding tasks with known security risks. In its first report, the generated code contained a security vulnerability in 45 percent of cases while being syntactically correct almost every time, and that rate barely moved through summer 2026. The October 2026 update shows the first meaningful improvement: the vulnerability rate fell to 38 percent. That is still more than one in three.
- Apps with open databases. In 2025, researchers showed that in apps built on the Lovable platform, the database's Row Level Security policies were insufficient, letting an unauthenticated attacker read or modify tables. The issue was recorded as CVE-2025-48757 and, according to reports, affected more than 170 projects.
- Moltbook. In early 2026, Wiz's security team showed that Moltbook, a social network for AI agents whose founder said he had not written a single line of its code himself, had its database connection key in client-side code with no database-level protection turned on. The result was read and write access to the entire production database, including around 1.5 million API tokens and tens of thousands of email addresses. It was fixed within hours of disclosure, but until then anyone who opened the page's code could reach the same data.
- An agent that wiped a database. In July 2025, Replit's coding agent, working on an experimental project, deleted the production database despite an explicit instruction to freeze changes. One reason was that, at the time, the development and live environments shared a single database. After the incident, Replit committed to automatically separating development and production databases.
What these incidents have in common is not AI itself. None of these mistakes is new: an unprotected database, keys in client-side code, and one environment shared by development and production are errors engineers have recognized for years. What has changed is speed and scale. These mistakes can now be built and shipped in a weekend, without anyone who would recognize them ever seeing the code.
Eight failure patterns in AI-built software
When we review an AI-built prototype for production, these are the first places we look:
| # | Pattern | What you see in the code | Consequence in production |
|---|---|---|---|
| 1 | Access control only in the UI | The button is hidden from regular users, but the API answers without checking; database access policies are off or far too open | Anyone with a little curiosity can reach everyone's data |
| 2 | Keys and secrets in the wrong place | API keys, database passwords or payment tokens in client-side code or in the repository | Account abuse, bills from paid services, data leaks |
| 3 | One environment for everything | Development, testing and production on one database and one server; the coding agent has production access | One experiment or wrong command damages real data |
| 4 | Tests for the happy path only | No tests, or tests the AI wrote for its own code that only check the expected case | Each new change breaks something elsewhere and nobody notices |
| 5 | Unaccounted-for dependencies | Many packages, outdated or unused; sometimes a package the tool suggested that may not be the genuine one | Known vulnerabilities and supply-chain risk |
| 6 | Data without migrations or backups | The database structure changed by hand or directly by the tool; no backup, or a restore that has never been tested | Lost data and no way back |
| 7 | Silent failures | Errors swallowed in the code; no structured logs, error tracking or alerts | Users report errors, not the system |
| 8 | Code nobody understands | Duplicated logic in several places, inconsistent structure, no documentation; the only "author" is the tool | Every change becomes slow and risky, and maintenance costs climb |
Two more usually stay hidden until the first busy day: performance under load, such as queries that run once per row or pages that load all the data at once; and missing rate limits on APIs backed by a paid service, including the language models themselves.
Rescue, partial rebuild or start over?
The first question a manager asks after an honest review is: "Do we keep this code or throw it away?" The answer is rarely either extreme. The goal is not to rewrite what the AI got right; it is to close, precisely, the gap between "it works on my machine" and "it runs safely in production".
| State of the codebase | Typical decision |
|---|---|
| The data model is sound, core flows work and the code is understandable after review | Continue on this codebase, hardening security, tests and the release path |
| Most of the system is healthy, but one sensitive part (authentication, payments, permissions) is built wrong | Rebuild that part with a proper design and keep the rest |
| The data model or the system's core boundaries are wrong, and security issues are everywhere | Build a new core, using the prototype as the product's living specification |
Even in the third case, the prototype is not worthless. A version real users have worked with is the most accurate requirements document a team can have, more accurate than anything written before building. We explain why data model and system boundary decisions matter so much in Software Architecture for Managers.
The path from prototype to production
Turning an AI-built prototype into a production system is not a vague, open-ended project. It has four stages, each with a clear output:
- Readiness assessment. An honest review of the architecture and codebase, a security and dependency review, and a list of risks in order of importance. The output is a decision to rescue, partially rebuild or start fresh, with an estimate based on the actual code, not guesswork.
- Hardening. Closing the high-risk holes before any new feature: access control on the server and in the database, moving secrets out of the code and rotating exposed keys, separating environments, and a tested backup.
- Safety net. Automated tests for business-critical paths, a CI/CD pipeline with a quality and security gate, and versioned database migrations. From here on, every change, whether a person or an AI wrote it, goes through this path.
- Operations and ownership. Error tracking, monitoring and alerting, release and rollback documentation, and a clear owner for the system after launch. Software that nobody is responsible for running is on its own at the first serious incident, however well it was built.
The order matters. Adding new features to a system that still lacks proper access control or backups only enlarges the surface that might leak tomorrow.
Don't drop AI; put a gate in front of it
None of this means teams should stop writing code with AI. The speed is real, and a team that ignores it falls behind. DORA's research describes AI as an amplifier: it magnifies whatever is already in a team's process, discipline or disorder. We cover this in more depth in DevOps in software development projects.
So the right question is: what gates does AI-generated code pass through before it reaches users? At Dropp, every pull request is reviewed by both a developer and AI. Before release, SonarQube checks code quality and Trivy checks for known vulnerabilities, and code that does not pass this gate does not ship, whether a person or a machine wrote it. We apply the same discipline through AI-built software productionization to software that was built outside this path.
Checklist before going to production
Before you put AI-built software in front of real users, answer these questions. If the answer to any of them is "I don't know", that is where the work starts.
- Has someone other than the tool read the code and taken responsibility for it?
- Is access control enforced on the server and in the database, not only in the UI?
- Are there no keys, passwords or tokens in client-side code or in the repository?
- Are development, testing and production separate, with no direct production access for AI tools?
- Do critical paths (sign-up, login, payments and anywhere user data is involved) have automated tests?
- Have dependencies been reviewed and scanned, and do you know what each one is for?
- Are database changes versioned, and has a backup restore been tested at least once?
- Are errors tracked, with alerts reaching the team before users notice?
- Has the system been tested under expected load, and do paid APIs have rate limits?
- Are the code repository, accounts and access in your organization's name?
- Is it clear who runs the system and responds to incidents after launch?
Conclusion
Vibe coding has cut the distance from idea to working version from months to days, and that has real value. But a working version is not yet a reliable product. What turns a demo into production is the same engineering discipline as ever: someone who understands the code, tests and a security gate in the release path, separate environments, monitoring, and an owner who answers for the system after launch. AI has not removed that discipline; it has made it matter more.
If you have a prototype built with AI that now has to hold up in front of real users, the Dropp engineering team can start with a production readiness review: we assess architecture, security, dependencies and tests, and make clear what should stay, what needs hardening and what should be rebuilt. More details are on our AI-built software productionization page.
Frequently asked questions
What is vibe coding in simple terms?
It is building software by telling an AI tool what you want and accepting the code it produces, without anyone reading or understanding that code. Andrej Karpathy coined the term in February 2025, and Collins named it its Word of the Year for 2025.
Is AI-generated code insecure?
Not inherently, but it cannot be trusted without review. Veracode's latest data shows generated code contained vulnerabilities in about 38 percent of security tests. That is why AI code needs the same review, testing and security scanning as any other code.
Is vibe coding a good way to build an MVP?
For testing an idea and showing it to users or investors, yes. But if the MVP will handle real user data or payments, at least access control, secrets management, environment separation and backups need to be in place before launch.
What is the difference between vibe coding and AI-assisted programming?
In vibe coding, nobody reads the code. In AI-assisted programming, an engineer uses AI for speed but reviews, understands and takes responsibility for every change. The difference is not how much AI is used; it is whether there is an owner and a quality gate.
Do we need to rewrite all the code the AI built?
Usually not. After a careful review, most prototypes either continue with hardened security, tests and a release path, or only their sensitive parts are rebuilt. Starting over makes sense when the data model or the system's core boundaries are wrong.
How long does it take to get a prototype to production?
It depends on the size and state of the codebase, and no reliable number can be given without seeing the code. That is why the right path starts with a readiness assessment, whose output is a list of risks, a decision about the codebase and a phased estimate.