Enterprise vibe coding: how to let non-engineers ship apps without leaking data

Enterprise vibe coding means employees who are not engineers build and publish internal tools with Lovable, Cursor, Claude Code or Replit. It works in production when four things hold: every table has access rules, no secret sits in the browser, one user cannot read another's records, and every app has a named owner. A short review before real data and a test of each live app after every change keep it that way.

The team behind Keybrum's attacksUpdated 9 min read

In short

  • The risk of vibe coding in a company is rarely the code's style. It is four concrete gaps: tables without access rules, secrets in the JavaScript, missing per-user checks, and apps that outlive their author.
  • A platform's own scanner helps and has limits its vendor states. Lovable's documentation says its security tools "cannot guarantee complete security" and recommends a further review for apps handling sensitive data.
  • A review gate of six questions, answered in fifteen minutes before an app touches real data, stops most leaks. Prototypes on fake data need no gate.
  • Split ownership in three: the builder owns the app, the data owner decides what data it may hold, and IT or security owns the inventory and the tests.
  • An app that passed review in March is a different app in June. Test what is live after every change, from outside, and close a finding only when the same test fails.
On this page

What enterprise vibe coding means

Vibe coding is building software by describing it to an AI tool and accepting what it writes. Enterprise vibe coding is the same practice inside a company, by people whose job is not engineering: an operations lead who builds a supplier portal, a finance analyst who builds an approval flow, a support manager who builds a refund dashboard. The tools are Lovable, Bolt, Replit and v0 for people who never open an editor, and Cursor or Claude Code for those who do.

The case for allowing it is simple. These people know the process better than anyone, and the tool they need would wait months in an engineering backlog. The case for rules is just as simple: the result is a real application on the internet, with a database, a login and API keys, and its author was never trained to ask who else can reach it.

If your first question is which of these apps already exist in your company, start with the inventory method in our shadow AI guide. This page is about the next step: how to say yes with rules that people follow.

Vibe coding in production: the four risks that matter

Public evidence on AI-built apps points at the same few gaps again and again. Veracode's 2025 GenAI Code Security Report ran 80 coding tasks through more than 100 language models and found a security flaw in 45% of cases. In the field, the research behind CVE-2025-48757 tested 1,645 apps built with Lovable and found missing or weak row level security in 170 of them. Broken access control is the first category of the OWASP Top 10:2025. In a company, four gaps do most of the damage.

RiskWhat goes wrongWhat a stranger can do
Data access rulesA table, a storage bucket or a Firestore collection has no rule, or a rule that lets everyone throughRead or change the data with the public key that ships in the app's JavaScript
KeysA secret key (an AI provider, Stripe, the Supabase service role) is pasted into frontend codeCopy it from the browser and use it: spend your credits, or bypass every database rule
Per-user accessThe app checks that you are signed in, and forgets to check that the record is yoursSign up, then read a colleague's or a customer's records by changing an id
Apps nobody ownsThe author changes teams or leaves. The app keeps running with its data and its keysNothing new, and that is the problem: the hole found next year has nobody to fix it

The first three have their own guides: Supabase RLS, API keys exposed in the frontend and IDOR and broken access control. The fourth is an organisation problem, and no scanner fixes it.

What the platforms already check, and where they stop

Use what the builders give you. Lovable is a good example of how far a platform goes today. Its documentation describes a Quick scan that runs when the publish dialog opens and checks database access rules, npm dependency vulnerabilities and whether an MCP server is exposed without authentication. A Deep scan, started by hand, reviews the code for access control, unsafe input, leaked secrets and exposed personal data. Workspace admins can turn on Block publishing with critical issues, which Lovable enables by default on new Enterprise workspaces it creates. A Security center on Business and Enterprise plans shows the state of every project.

Turn all of that on. Then read the limit Lovable states itself: "These tools help identify common security issues, but they cannot guarantee complete security." The same page recommends an additional professional review for apps that handle sensitive data. That is honest, and it is true of every builder. A platform's scanner reads the project it generated. It does not attack the deployed app from outside, as two different users, on the day after someone changed a rule. The Lovable security guide goes through what remains.

A lightweight review gate before real data

Do not review prototypes. A demo on invented data can be built and thrown away freely. The gate applies at one moment: before the app holds real customer, employee or financial data, or before anyone outside the author's team gets the link. At that moment, the author and one reviewer answer six questions. It takes about fifteen minutes.

  • Who owns this app? One named person, and a second who can reach the hosting account if the first is away.
  • What data does it hold? List the tables. Anything the data owner has not approved comes out.
  • Who can sign up and sign in? Company accounts only, unless customers are meant to use it.
  • Does every table, bucket or collection have an access rule? On Supabase, the query below must return no rows.
  • Is any secret in the frontend? Search the deployed JavaScript for sk_, service_role and -----BEGIN. Or paste that JavaScript into the API key leak checker.
  • Was the live app tested from outside, and where is the proof? A dated report with what was tried and what worked.
sql
select schemaname, tablename
from pg_tables
where schemaname = 'public'
  and rowsecurity = false;
Supabase SQL editor: tables in the public schema with row level security off. The gate passes when this returns nothing.

Row level security turned on is the floor. A policy that says using (true) is on and lets everyone through. That is why the sixth question exists: only a request from outside shows what the rules really allow. Write the answers in the inventory next to the app. The vibe coding policy template turns this gate into rules you can copy.

Who owns what

Most failures here are gaps between people. Each assumed another one had checked. Three roles are enough for a company of any size.

RoleOwnsDoes not own
The builder (the employee who made the app)The app itself: fixing findings, keeping it working, declaring changes, retiring itThe decision on which data the app may hold
The data owner (head of the team whose data it is)The yes or no on each class of data, and who may see itThe technical checks
IT or securityThe list of allowed tools, the inventory, the review gate, the tests on what is liveBuilding or fixing the apps

Two rules keep this from falling apart. Every app lives in a company account, never a personal one, so access survives a departure. And when a builder leaves or changes teams, their apps are handed over or retired in the same week as their laptop.

Test what is live, continuously

A review is a photo of one day. AI-built apps change faster than any other software in the company, because changing them costs one prompt. A new feature adds a table. A fix for a login error loosens a rule. The agent that wrote the app is the same one that reviews it, and it stops when the app works.

So the last part of the method is a test that repeats. It runs against the deployed app, from the internet, with what any visitor gets. It runs after each publish and on a schedule. And a finding is closed when the same request that worked now fails, and no sooner. The vibe coding security checklist lists the eight layers to cover in each run.

  1. After every publish: the six gate questions that can be automated (rules, keys, sign-up, per-user access).
  2. Every week or month: the full test, because the attack methods change even when the app does not.
  3. Every quarter: the inventory itself. New apps appear, owners move.

What Keybrum covers here, and what it does not

Keybrum is the automated attack test we build for apps made with AI tools. For a company, the Team plan does three parts of this method. It finds every app deployed in the Vercel, Netlify and Supabase accounts you connect, including undeclared ones, and ranks them by what they expose, each with an owner. It attacks each app whose ownership is proven from outside: tables, storage and database functions read with the public key, provider keys in the JavaScript, one account reading or changing another account's records, privileged fields set from the client, admin pages open to anyone and AI routes that answer with no session. And it attacks again on every deploy and every week or month.

Each finding carries the request that worked and a fix prompt for Lovable, Bolt, Cursor or Claude Code, so the builder can close it without a security background. The Team plan is €99 a month excluding VAT for 25 apps and 10 members, with your brand on the reports. For one app, a one-time report is €49. Above 25 apps there is no listed plan: contact us.

Questions

Is vibe coding safe for production?

It can be, when the app is tested as it is deployed. The common failures are known: missing access rules, secrets in the frontend, missing per-user checks. Each can be checked on the live app in minutes. What is unsafe is publishing on real data because the app works, with no test from outside.

Should non-engineers be allowed to build internal tools with AI?

Yes, with a named owner, approved data and a review before real data goes in. A ban tends to move the same work to personal accounts where nobody sees it. A short gate and a repeated test give the company the tools and a view of the risk.

What should a vibe coding review cover?

Six points: a named owner, the data the app holds, who can sign up, an access rule on every table or bucket, no secret in the frontend, and a dated test of the live app from outside. Prototypes on invented data do not need it.

Who is responsible when an employee-built app leaks data?

Towards customers and regulators, the company answers for it. Inside the company, it should be clear before the leak: the builder owns the app and its fixes, the data owner approved what it holds, and IT or security owns the inventory and the tests. If none of the three knew the app existed, the gap is the inventory.

Do the security scanners built into AI builders make a separate test unnecessary?

No. They catch common problems early, and you should turn them on. Lovable's own documentation says its tools cannot promise complete security and recommends a further review for sensitive apps. A built-in scanner reads the project. A test from outside shows what a stranger can do to the deployed app.

How is enterprise vibe coding different from low-code?

Low-code platforms keep the app inside one vendor's runtime, with its permission model. AI builders generate ordinary code and a real backend, often Supabase or Firebase, that the author controls fully. That gives more freedom and moves the access rules, the keys and the hosting into the hands of someone who may never have set them before.

Sources

  1. 1.Security overview, Lovable Documentation
  2. 2.AI-generated code poses major security risks in nearly half of all development tasks (2025 GenAI Code Security Report), Veracode
  3. 3.CVE-2025-48757, NIST National Vulnerability Database
  4. 4.Statement on CVE-2025-48757, Matt Palmer
  5. 5.Row Level Security, Supabase Docs
  6. 6.OWASP Top 10:2025, OWASP

Your app is public. So are its mistakes.

Check what any attacker can already see in it, free, in ten seconds. The full attack, with proof and fixes: €49 once, or €29 a month on a plan.

or continue with Google or GitHub