What is vibe coding? And where vibe-coded apps break

Vibe coding is building software by describing what you want to an AI tool in plain English and accepting the code it writes, often without reading it. The term dates from February 2025. It is fast for prototypes, but vibe-coded apps tend to break at the same points: security, data, payments, search visibility and real traffic.

Vibe coding went from a single social media post to a dictionary award in nine months. The computer scientist Andrej Karpathy coined the term in February 2025, and in November 2025 Collins named vibe coding its Word of the Year 2025, defining it as "the use of artificial intelligence prompted by natural language to write computer code". Tools such as Lovable, Bolt, Replit, Cursor, Claude Code and Codex now let a founder, a marketer or an operations lead get a working app on screen in an afternoon.

That is genuinely useful. It also means a lot of software is going live that nobody has read line by line. This guide is practical rather than promotional: what vibe coding is, how widely it is used, what the research says about speed and security, and the specific places where vibe-coded apps fail once real people start using them.

  • Vibe coding means describing software in plain English and letting an AI write the code, judged by whether the app works.
  • It is a spectrum: the NCSC advises matching the level of oversight to what the code does and touches.
  • 84% of developers use or plan to use AI tools, yet 72% say vibe coding is not part of their professional work (Stack Overflow, 2025).
  • In Veracode's 2025 tests, 45% of AI-generated code samples failed security tests.
  • Most vibe-coded apps can be hardened and kept; a full rebuild is rarely the first answer.
Key takeaways

What does vibe coding mean?

Vibe coding is a way of building software where you steer an AI model with natural language and judge the result by whether it works, not by reading the code. You describe a screen, a feature or a fix; the tool writes and changes the code; you look at the preview and ask for the next change.

In practice it is a spectrum rather than one technique. At one end, someone with no programming background builds a whole app in Lovable or Bolt and never opens a file. At the other, an experienced developer uses Claude Code, Codex or Cursor to write most of the code but reviews every change before it is merged. The UK's National Cyber Security Centre (NCSC) makes the same point in its June 2026 blog on the vibe coding spectrum: "Different code deserves different levels of oversight."

ApproachWho reads the codeGood for
Pure vibe codingNobody; the preview is the testPrototypes, pitch demos, throwaway experiments
Reviewed vibe codingA developer checks the risky parts before launchInternal tools and early products with real users
AI-assisted engineeringEvery change is reviewed, tested and versionedCustomer portals, payments, personal data
The vibe coding spectrum

Where did the term vibe coding come from?

The phrase comes from a post Andrej Karpathy, a co-founder of OpenAI, published on X in February 2025, describing a style of programming where you "fully give in to the vibes" and stop reading the code. Within weeks, Y Combinator managing partner Jared Friedman said that "a quarter of the W25 startup batch have 95% of their codebases generated by AI", as reported by TechCrunch in March 2025.

Vibe coding in numbers
  • 2025 the year Collins made vibe coding its Word of the Year
  • 25% of Y Combinator's Winter 2025 batch had 95% AI-generated codebases
  • 84% of developers use or plan to use AI tools in their development process

Sources: Collins, Word of the Year 2025; TechCrunch, Y Combinator W25, 2025; Stack Overflow, Developer Survey 2025

How does vibe coding work in practice?

It works as a short loop of prompt, preview and refine, and every AI app builder follows it, whichever brand is on the screen:

  1. Describe the app or the change in plain English
  2. The AI writes or edits the code and the database
  3. Check the live preview and click through it
  4. Ask for fixes and new features, one at a time
  5. Publish to a hosted address or your own domain
The vibe coding loop

Browser-based builders (Lovable, Bolt, Replit) handle hosting, a database and sign-in for you, which is why they feel so quick. Coding agents (Claude Code from Anthropic, Codex from OpenAI, and Cursor) work inside a real codebase on your own machine or in the cloud, which gives you more control and more responsibility. We compare the builders side by side in Lovable vs Bolt vs Replit vs Cursor, explain how teams use the agents in Claude Code for business, and cover the platform itself on our Lovable developer page.

How many developers actually vibe code?

Fewer than the hype suggests. Almost every developer now uses AI tools, but most say they review what the tools produce rather than vibe coding. In the Stack Overflow Developer Survey 2025, 84% of respondents were using or planning to use AI tools and 51% of professional developers used them daily, yet 72% said vibe coding was not part of their professional work. Trust is the sticking point: more developers distrust the accuracy of AI output than trust it.

How developers use and rate AI coding tools

Daily use is among professional developers; other rows are all respondents.

Survey answerShare
Use or plan to use AI tools84%
Not vibe coding at work72%
AI answers almost right66%
Use AI tools daily51%
Distrust AI accuracy46%
Trust AI accuracy33%

Source: Stack Overflow, Developer Survey 2025: AI

The biggest single frustration, named by 66% of developers, was "AI solutions that are almost right, but not quite." That is the vibe coding problem: code that passes a glance at the preview and fails on the case nobody tried.

How widely are UK businesses using AI?

About three in ten UK businesses now use at least one AI technology, and the share is far higher in larger firms. The Office for National Statistics (ONS) found that use among businesses with 10 or more employees rose from around 12% in late 2023 to around 35% in June 2026, according to its July 2026 analysis AI in UK businesses: 2023 to 2026. Its October 2026 business survey bulletin puts the figure at 30% of all businesses in September 2026.

UK businesses using at least one AI technology
Business size and dateShare
10+ staff, late 202312%
0 to 9 staff, June 202628%
10+ staff, June 202635%
250+ staff, June 202649%

Source: ONS, Artificial intelligence in UK businesses: 2023 to 2026, July 2026

The ONS does not ask about vibe coding specifically, but it does show that large language models were the most widely used AI type among firms with 10 or more employees in June 2026, at 18%. Those are the models behind every vibe coding tool.

Does vibe coding make you faster?

For prototypes, usually yes; for experienced developers on mature code, the evidence is mixed. The strongest study is a 2025 randomised controlled trial by the research group METR, which gave 16 experienced open-source developers 246 real tasks and randomly allowed or disallowed AI tools on each one.

What METR measured in early 2025
  • 24% speed-up the developers expected before they started
  • 20% speed-up they still believed they had got afterwards
  • 19% longer the tasks actually took when AI tools were allowed

Source: METR, Measuring the impact of early-2025 AI on experienced open-source developer productivity, 2025

The gap between how fast it felt and how fast it was is the useful lesson. METR's February 2026 update says it is "likely that developers are more sped up from AI tools now", but calls its newer data "only very weak evidence for the size of this increase." For a business, vibe coding is fastest at the start, when there is little code to break, and slows as the app grows unless someone keeps the structure tidy.

Is vibe coding safe?

Vibe coding is safe for experiments and unsafe by default for anything that holds personal data or takes money, unless someone checks the parts that matter. The AI is good at making features work. It is less reliable at making them work only for the right people.

The research backs this up. Veracode's 2025 GenAI Code Security Report tested code from more than 100 large language models across Java, JavaScript, Python and C#, and found that "45% of code samples failed security tests and introduced OWASP Top 10 security vulnerabilities into the code."

AI-generated code samples that failed security tests

Code from over 100 large language models, tested by Veracode in 2025.

Test setFailure rate
All code samples45%
Java samples72%
Cross-site scripting cases86%

Source: Veracode, 2025 GenAI Code Security Report

The platforms are candid about this. Lovable runs a quick security scan from the publish dialog, and its security documentation says the tools "cannot guarantee complete security." Replit offers Agent security scans to all paid builders and states plainly that "a scan is not a complete security review." Those scans are worth running; they are a floor, not a sign-off. The NCSC puts it well: "The risk isn't in using AI. The risk is not applying the right safeguards when the stakes are high." A professional vibe-coded app security audit is the step between the two.

The other half of safety is legal. If your app stores customer data, UK GDPR requires you to process personal data securely by means of "appropriate technical and organisational measures", as the Information Commissioner's Office (ICO) guide to data security explains, whoever or whatever wrote the code. The threat is not theoretical: the government's Cyber Security Breaches Survey 2025 found that 43% of UK businesses experienced a cyber security breach or attack in the previous 12 months.

Where do vibe-coded apps break?

Vibe-coded apps rarely fail on the demo path. They fail on the paths nobody asked the AI to think about: another user's data, an expired session, a refund, a search engine, a spike in sign-ups. These are the areas to check first.

AreaWhat tends to go wrongWhat to check
Access controlDatabase tables readable or writable by anyone with the public keyRow Level Security on every exposed table, with policies tested as two different users
SecretsSecret or service keys shipped in browser code or committed to GitOnly publishable keys in the front end; secrets in server functions or environment variables
AuthenticationPassword reset, email confirmation and sessions that work in the preview but not in productionYour own email (SMTP) sender, redirect URLs on the live domain, rate limits
PaymentsOrders marked paid by the browser rather than by the payment providerWebhooks verified on the server; prices set server-side, never trusted from the client
Data modelDuplicated tables and columns after many rounds of promptingA clear schema, migrations in version control, backups you have restored at least once
Search visibilityPages rendered only in the browser, generic titles, no sitemapServer-rendered or pre-rendered HTML, unique titles and descriptions, a sitemap
Scale and costSlow queries and runaway usage bills once real traffic arrivesIndexes on the columns you filter by, pagination, usage alerts
MaintainabilityEvery new prompt breaks an older featureCode in Git, a few automated tests on the critical paths, one person who understands the structure
Common failure points in vibe-coded apps

Access control is the one to fix first. Supabase, which sits behind Lovable's built-in backend and many other vibe-coded apps, is explicit: enable Row Level Security on every table in an exposed schema, and never put a secret key in the browser. Our vibe-coded app security checklist walks through these checks one by one.

Search visibility is the one people discover last. Older Lovable apps, built on React and Vite, render in the browser, and Lovable serves pre-rendered copies of their published pages to verified search engines. According to the Lovable FAQ, new apps created from 13 May 2026 (22 June 2026 for Enterprise workspaces) use TanStack Start, which renders pages on the server. Either way, titles, descriptions, internal links and content still need deliberate work. We cover hosting, domains and SEO in taking a Lovable app to production.

When is vibe coding the right choice?

Vibe coding is the right choice when speed of learning matters more than durability. Good fits include:

  • Prototypes and pitch demos that need to look and feel real for a few weeks.
  • Internal tools used by a small, trusted team, such as a quoting calculator or a stock tracker.
  • Demand tests: a landing page and sign-up flow to see whether anyone wants the product before you fund a full build.
  • A head start on a real product, as long as you budget for a review and hardening phase before launch.

It is a poor fit, without professional review, for apps that handle health information, children's data, financial records or payments at volume, or anything where a data leak would have to be reported to the ICO. If the idea has proved itself and needs to become a dependable first version, that is the point at which app MVP development with proper review takes over from pure prompting.

Should you fix a vibe-coded app or rebuild it?

Fix it first, in most cases. A vibe-coded app usually contains months of product thinking: the screens, the flows, the wording your users already understand. The pragmatic route is to move the code into version control, close the security gaps, untangle the data model, add tests around the critical paths and then keep building, with or without the AI tool.

A rebuild makes sense when the data model cannot support the next stage of the business, when the app has to meet a standard the platform cannot reach, or when the code has been regenerated so many times that no one can predict what a change will break.

Circlemount is based at St James House in Telford, Shropshire, and Marcus (founder) and a team of 12 specialists fix vibe-coded apps for businesses across the UK. If you are nearby, see our pages for Telford and Birmingham; the same service is available to teams in Manchester and London.

What should you check before you launch?

Before real customers use a vibe-coded app, work through this short list:

  1. Run the platform's own security scan and fix everything it flags.
  2. Confirm Row Level Security is on for every table, and test it while signed in as two different users.
  3. Search the code for secret keys and move any you find to the server.
  4. Set up your own email sender, and test sign-up, login and password reset on the live domain.
  5. Connect the project to GitHub (or GitLab or Bitbucket) so you own a copy of the code.
  6. Turn on backups and usage alerts.

If you are deciding which tool to use for the next build, or how to turn a prototype into something you can maintain, our guides to agentic AI and Codex vs Claude Code explain the coding agents that sit behind much of this work, and our vibe-coded app services cover everything from a first review to a finished product.

Frequently asked questions

Who coined the term vibe coding?

The computer scientist Andrej Karpathy, a co-founder of OpenAI, coined the term in a post on X in February 2025. He described giving in to the vibes and letting the AI write the code. Collins named vibe coding its Word of the Year for 2025, which shows how quickly the idea spread beyond developers.

Is vibe coding safe for a business app?

It can be, once the risky parts have been checked by someone who reads code. The common gaps are database access rules, exposed secret keys, authentication and payments. The NCSC advises more rigour where stakes are high, and UK GDPR still requires appropriate security for any personal data the app stores.

Does vibe coding make developers faster?

It depends on the work. A 2025 METR trial found experienced open-source developers took 19% longer with AI tools, while believing they were faster. METR's 2026 update suggests newer tools probably help more. For a prototype with little existing code, prompting is usually much quicker than writing everything by hand.

Can a vibe-coded app be fixed, or does it need a rebuild?

Most can be fixed. The usual route is to put the code in version control, close the security gaps, tidy the data model and add tests to the critical paths. A rebuild is worth considering only when the data model cannot support the business or nobody can predict what a change will break.

Do I own the code in a vibe-coded app?

Check each platform's terms, but most let you export or sync your code. Lovable offers two-way sync with GitHub and also supports GitLab and Bitbucket, and coding agents such as Claude Code and Codex work directly in your own repository. Keeping a copy in your own Git account is the simplest protection.

Is vibe coding the same as AI-assisted coding?

Not quite. Vibe coding usually means accepting AI-written code without reading it, judged only by whether the app works. AI-assisted coding covers the same tools used by developers who review, test and understand each change. The tools are the same; the level of oversight is what differs.

Will a vibe-coded app rank on Google?

It can, but ranking is not automatic. Search engines need readable HTML, unique titles and descriptions, sensible internal links and useful content. Newer Lovable apps render on the server, and older ones serve pre-rendered pages to verified crawlers, but the content and structure still have to be planned.

Sources

Checked on 16 November 2026. Product features and figures change quickly, so check the current source before you rely on a detail.

Need a vibe-coded app fixed or finished?

Circlemount fixes, secures and finishes apps built with Lovable, Bolt, Replit, Cursor and Claude Code, so they are ready for real customers. See how we fix vibe-coded apps, browse our vibe-coded app services, or Join the Circle for a free business analysis and proposal.