Skip to content

Is vibe coding bad?

By Mari Luukkainen · 8 min read · Last updated

On this page

Vibe coding is not bad, but it has four real risks: security holes, code nobody understands, costs that grow with use, and skills you never learn. Each can be reduced with specific checks. For prototypes and personal tools the risk is low. For anything that holds other people's data, it is high.

Below, each risk comes with what it looks like, a public example where one exists, and a check you can run today. A project that handles payments or health data should be reviewed by a person who can read code. If you are new to the term, what is vibe coding explains it.

Ask AI about this pageChatGPTClaudePerplexityGoogle AI ModeGrok

Is vibe coding a security risk?

Yes. An AI writes code that works on the screen you tested and says nothing about the screens you did not test. Two public cases show how it goes wrong.

  • In July 2025 SaaStr founder Jason Lemkin tested Replit's AI agent and it deleted the production database, which held records on 1,200+ executives and 1,190+ companies. The agent first said a rollback would not work. Lemkin recovered the data, and Replit's CEO called the incident "unacceptable and should never be possible" (Fortune).
  • CVE-2025-48757, published 30 May 2025 with a CVSS score of 9.3, describes an insufficient Row-Level Security policy in Lovable through 15 April 2025. It let unauthenticated attackers read or write arbitrary database tables of generated sites (NVD).

The usual holes in vibe coded projects are specific and repeat:

  • Secret keys in the wrong place. An API key in code that runs in the browser can be read by any visitor. Keys belong on the server, in environment settings, never in Git.
  • Missing permission checks. A page shows only your orders because the link has your ID, but changing the ID in the address shows someone else's order.
  • Database tables with no access rules. On Supabase, Row Level Security (RLS) must be on for every table, and each table needs a policy that limits rows to their owner. With RLS off, anyone holding the public key can read the table.
  • No limits on forms. A contact form or login that accepts unlimited requests can be used to send spam or guess passwords.
  • Trusting what the user types. Text from a form placed straight into a database query or onto a page can break the app or run an attacker's code.
  • Invented or abandoned packages. The AI sometimes names a package that does not exist. Check that a package is real and maintained before you install it.

What to do

Run a security pass as its own step, in a fresh session, so the tool is not defending its own work:

security prompt
Review this project for security problems. Do not change code yet. List, with file names:
1. Any secret key, token or password that appears in the code or would be sent to the browser.
2. Any page or API route that returns data without checking who is asking.
3. Any database table without Row Level Security, or with a policy that lets one user read another user's rows.
4. Any form or API route with no rate limit.
5. Any place where user input goes into a database query, a file path or HTML without validation.
Then rate each finding high, medium or low and propose a fix.

Then check three things by hand. Search the project for "key" and "secret" and look at what appears. Log in as one test user and try to open another test user's page by editing the URL. Confirm that .env files are listed in .gitignore. Keep real customer data out of any project that has not been reviewed, and give agents a test database, never the production one.

Vibe Coding Has A Security Problem (And How To Fix It) · Chris Raroque

Is vibe coded software maintainable?

Only if something keeps it tidy. The AI adds code in the direction of your latest prompt. After forty prompts with no structure, the same logic appears in three places, one of them slightly different, and a change in one place breaks another.

  • You ask for a small change and the tool rewrites four files.
  • A bug you fixed last week returns after an unrelated change.
  • Nobody, including you, can say which file controls the price shown at checkout.

What to do

  • Write the rules once in the tool's instruction file: the stack, the folder layout, naming, and what must not change. AGENTS.md is read by Codex, Claude Code, GitHub Copilot, OpenCode and Cursor.
  • Commit after every working round and name the commit after what changed.
  • Ask for a structure review every ten rounds or so.
structure prompt
Review the structure of this project. Do not change code. List duplicated logic, files over 300 lines, and any function whose name no longer matches what it does. Propose a refactor plan in small steps, each one safe to commit on its own.

I run more than 100 apps and barely read the code anymore (post). What keeps that workable is a private handbook of decisions that the AI reads before it starts. Every bad day on my setup came from configuration drift that nothing was watching. The worst was a hand-edit to a compose file that a tool then rewrote, which silently deleted a production database container's definition and left it one flaky deploy away from being erased. The fix was to make the system refuse to run when reality does not match what it expects: validate the file before touching it, and alert when a backup has not succeeded in a day (post).

Does vibe coding cost more than it seems?

The first version is cheap. The later costs are usage-based: AI tool usage grows with how much you prompt, and hosting, databases, email and any AI feature inside your app grow with how many people use it.

CostWhen it appearsCheck
AI tool usageMany long sessions on a large projectRead the tool's usage page weekly and set the limits it offers
Hosting and databaseWhen real visitors arriveLook up the host's plan limits before launch, not after
AI calls inside your appEvery user action that calls a modelSet a spending cap with the provider and cache repeated answers
ReworkWhen the structure has driftedDo the structure review above before adding features

My numbers: the 100+ apps run on 9 Hetzner boxes for about 200 EUR a month, almost all of it Hetzner. The smallest box is 5 EUR, the biggest has 16 GB and runs 17 apps, and the total includes my own portfolio and Herizon. My estimate for the same setup on per-service flat fees plus error tracking, uptime monitoring and a CI quota is 1,500 to 2,000 EUR a month, maybe more. A like-for-like comparison with Vercel or Render is not honest, because most of my apps have too little traffic to produce real variable cost (post). Prices for the AI tools themselves are in best vibe coding tools.

Do you learn anything from vibe coding?

You learn what you practise. If you accept every change without reading, you learn how to describe software and how to test it, and little about code. That is a real skill, but it leaves you stuck when the AI cannot fix a problem.

I avoided the modern JavaScript stack for over a decade because LAMP worked for my marketing funnels. With Cursor explaining things and ChatGPT breaking down the differences, I learned it in four hours. I still do not understand everything, and I wrote in March 2025 that learning while building is the process (post).

learning prompt
Explain the change you just made in plain language. Which file did you change and why? What would break if I removed it? Give me one thing in this change that I should be able to explain to someone else.

Once a week, pick one file and ask the tool to quiz you on it. The aim is to know enough to tell a good fix from a bad one.

When is vibe coding fine and when is it not?

ProjectRiskApproach
Personal tool, no outside usersLowVibe code freely, commit often
Prototype to test an idea with a few peopleLow to mediumNo real personal data, throw it away if it works
Public site with a contact formMediumSecurity prompt, rate limits, spam protection
App with accounts and saved user dataHighSecurity pass, RLS and permission tests, review by someone who reads code
Payments, health, children's dataVery highProfessional review before launch, managed payment and login services

Simon Willison separates vibe coding, where you do not review the code, from AI-assisted programming, where you do (post). The rows above are about how much review a project needs.

Vibe coding in prod | Code w/ Claude · Anthropic

A short checklist before you share a project

  1. Search the code and Git history for keys, tokens and passwords.
  2. Run the security prompt in a fresh session and fix the high findings.
  3. Open another test user's page by editing the URL. It must fail.
  4. On Supabase, confirm RLS is on for every table.
  5. Run npm run build and fix every error without turning checks off.
  6. Write the monthly cost estimate using the vendors' current prices.
  7. Write down the three files a new person would open first, so you can find them too.

Next: how to vibe code for a full first project, and the prompt library for security and test prompts.

Frequently asked questions

Does AI make experienced developers slower?

METR's July 2025 randomized trial with 16 experienced open source developers on 246 issues found they took 19% longer with AI, while expecting a 24% speedup. In February 2026 METR said its late-2025 data was an unreliable signal, with returning developers estimated at an 18% time reduction and new ones at 4%, both with wide uncertainty, and that developers were probably more sped up in early 2026.

How often does AI-generated code have security flaws?

Veracode's July 2025 report on more than 100 language models found that 45% of code samples failed security tests and introduced OWASP Top 10 vulnerabilities. Java failed in 72% of tasks, and cross-site scripting defences failed in 86%. It is vendor research.

Who is responsible for a security hole in a vibe coded app: the builder or me?

Assume you are. Lovable disputes CVE-2025-48757 and says each customer accepts responsibility for their own project's configuration. Test your own access rules whichever tool you use.

How do I stop an AI agent from touching my production data?

Give the agent credentials for a test database only, never the production keys, and keep production behind a separate login you use yourself. In the Replit case the agent first said a rollback would not work, so backups need to be somewhere the agent cannot reach.

Next guideVibe coding jobs →