Skip to content

Is your vibe-coded app safe to put online?

20 security checks for apps built with Lovable, Bolt, Cursor or v0: row-level security, keys kept out of the browser, owner checks, rate limits and AI risks.

Free. Runs in your browser. No sign-up.

0 of 20 done

Tick what your app already does. Your ticks are saved in this browser only.

Who can read your data

0 of 5 done

  • With Supabase, the browser talks to your database directly using a public key. Any table without row-level security can be read, and often changed, by anyone who opens the network tab.

    Ask your coding agent

    List every table in our database that the browser can reach (in Supabase, every table in the public schema) and whether row-level security is on. For each table without it, write a migration that turns it on and adds the narrowest policies the app needs: users read and change only their own rows, matched on auth.uid(), and tables the browser never touches get no policies at all. Show me the policies before applying them, and list anything in the app that will stop working. If we use Firebase, do the same with its security rules.

  • A policy that just says true, or lets the anon role in, turns row-level security on in name only. Everyone it applies to can still read every row.

    Ask your coding agent

    Review every row-level security policy in our database, or every Firebase security rule. Flag any whose condition is just true, that lets the anon or public role read or write without a clear reason, or that doesn't compare the row's owner with the signed-in user (auth.uid() in Supabase). For each one, explain what it allows today, propose a stricter version and write the migration.

  • The usual leak is one signed-in user seeing another's data, which testing with a single account never shows. Changing an id in the URL is often all it takes.

    Ask your coding agent

    Help me prove that one user can't see or change another user's data. Create two test users, A and B, each with some data. Signed in as B, try to read, update and delete A's records through every API route, server function and database query the app makes, including by changing ids in URLs and request bodies. Report each attempt and whether it was blocked, then fix the ones that weren't.

  • Supabase's service_role or secret key, and Firebase admin credentials, skip every security rule. If one reaches the browser, anyone can read and change your whole database.

    Ask your coding agent

    Search the code, env files and build settings for the Supabase service_role or secret key, Firebase admin credentials and any other database key that bypasses security rules. Confirm each one is used only in server code, and never in a variable with a public prefix such as NEXT_PUBLIC_, VITE_ or EXPO_PUBLIC_. Move any that aren't, and tell me which were exposed so I can rotate them.

  • User uploads often land in a public storage bucket, where anyone with the link can download them. Links get shared, logged and guessed.

    Ask your coding agent

    List every storage bucket the app uses (Supabase Storage, Firebase Storage, S3 or similar) and whether it's public. For buckets that hold user uploads, make them private, add policies so each user can reach only their own files, and serve files through signed URLs that expire. Tell me which buckets were public and what they held.

Keys and secrets

0 of 4 done

  • Anything in the browser bundle can be copied, including env variables with a public prefix. A leaked AI key can empty your credit balance in a day.

    Ask your coding agent

    Search the source code and the built browser bundle for secret keys: AI providers (OpenAI, Anthropic, Google, OpenRouter, ElevenLabs, Replicate, fal and similar), payments, email, SMS and database admin keys. Check every env variable with a public prefix such as NEXT_PUBLIC_, VITE_ or EXPO_PUBLIC_. Move each secret into a server-only variable used from a server route or function, and list every key that was exposed so I can rotate it.

  • A key committed once stays in the history after you delete it, and bots scan public repos for keys. Local settings files, such as Claude Code's settings.local.json, have leaked keys this way.

    Ask your coding agent

    Check that no secrets are committed to this repository, including in past commits. Make sure .env files, local settings files (such as .claude/settings.local.json) and credential files are listed in .gitignore. Scan the full history with gitleaks or a similar scanner and list every secret it finds, with the file and commit. Don't rewrite history yet; tell me which keys to revoke first.

  • Deleting a leaked key from your code doesn't stop anyone who already copied it. Only revoking it at the provider does.

    Ask your coding agent

    Make a table of every API key and secret this app uses: which service issues it, every place it's set (local env, hosting provider, Supabase or Firebase secrets, CI), and whether it has ever been in the browser bundle, a git commit, a screenshot or a chat. Don't change any code. I'll revoke and replace the exposed ones, then help me update each place the new key goes.

  • If a key leaks anyway, a limit at the provider caps the damage. Some providers only send alerts, so check which ones actually stop requests.

    Ask your coding agent

    List every paid API this app calls (AI, email, SMS, maps and similar). For each, tell me whether its dashboard offers a hard spending limit or only alerts, where to set it, and a sensible limit based on how our code uses it. Don't change any code.

Every request checks who's asking

0 of 5 done

  • Hiding a button doesn't protect the route behind it. Anyone can call your API routes and server functions directly.

    Ask your coding agent

    List every API route, server action, edge function and webhook in this app. For each, tell me whether it checks on the server that the caller is signed in and allowed to do this. Add the missing checks, returning 401 or 403 when they fail. Webhooks should verify the sender's signature instead. List what you changed.

  • Being signed in isn't enough. A route that loads an order or a document by id must check it belongs to the caller, or changing the id shows someone else's.

    Ask your coding agent

    Find every server route or function that reads, updates or deletes a record using an id from the request. Check that each one confirms the record belongs to the signed-in user, or to their team, before using it. Fix any that don't, and list each route and what you changed.

  • An admin page left out of the menu can still be opened by URL, and a role kept in the browser can be edited.

    Ask your coding agent

    Find every admin or staff-only page, route and function. Check that each one verifies the user's role on the server, from the database or the auth provider, not from local storage, an editable cookie or a URL parameter. Fix any that don't, and list them.

  • If the browser sends the price, the credit balance or the user's plan, anyone can change it in the network tab before it reaches you.

    Ask your coding agent

    Find every request where the browser sends a price, amount, credit balance, plan, role or discount. Make the server look each of these up itself, from the database or the payment provider, instead of trusting the request. List each case and what you changed.

  • Checks in a form only stop honest users. The server has to reject wrong types, oversized inputs and unexpected fields itself.

    Ask your coding agent

    Add server-side validation to every API route and server function that takes input, using a schema library such as zod. Reject unknown fields, cap the size of text and files, and return a clear 400 error. Confirm every database query is parameterised and none is built by joining strings.

Abuse and surprise bills

0 of 3 done

  • Without them, bots can guess passwords, create thousands of free accounts and hammer every endpoint that costs you money.

    Ask your coding agent

    Add rate limits to sign-up, sign-in, password reset and every route that calls a paid API (AI, email, SMS). Limit by IP address and, when signed in, by user id. Keep the counters somewhere shared across server instances, not in memory, and return HTTP 429 with how long to wait. List the routes you protected.

  • Builders report scripts creating accounts to farm free AI credits, paid for by the builder. A verified email and a bot check at sign-up slow that down.

    Ask your coding agent

    Make our free tier harder to farm. Grant free credits only after the email address is verified, block disposable email domains, add a bot check such as Cloudflare Turnstile to sign-up, and limit new free accounts per IP address per day. Tell me which of these we already have, and what each change means for a real user signing up.

  • One account, or a stolen login, can call your AI feature thousands of times overnight. A cap per user turns that into a polite error instead of a bill.

    Ask your coding agent

    Add a per-user daily spending cap for AI calls. Before every AI provider call on the server, add up that user's AI cost for today from the database. If it's over the cap (one config constant, start at $2), don't call the provider: return a clear error saying the daily limit was reached and when it resets, and show that message in the UI.

AI features

0 of 3 done

  • Text the AI reads, from a user, a web page or an uploaded file, can tell it to ignore your instructions. Limit what each AI step can reach, so a hijacked one can't do damage.

    Ask your coding agent

    List every AI step in this app that can call tools, run code, read files, send messages or query the database. Make sure each one acts only with the signed-in user's own permissions, can't reach other users' data or any secrets, and asks the user to confirm anything that can't be undone, such as sending, deleting or paying. Tell me which steps could do more than that, and what you changed.

  • Users can often get a model to repeat its instructions. Treat anything in a system prompt, such as keys, internal links or other customers' data, as public.

    Ask your coding agent

    Check every system prompt and prompt template in this app for API keys, passwords, internal URLs, other users' data or anything we wouldn't show a user. Move secrets out, and pass the model only data the current user is allowed to see. List what you found.

  • A model's answer can carry HTML, links or code planted in its input. Showing it raw, or running it, passes the attack on to your user.

    Ask your coding agent

    Find every place AI output is shown in the UI, saved, or used to build a query, URL, file path or command. Make sure it's escaped or sanitised before it's shown (no dangerouslySetInnerHTML or v-html on raw model output; render markdown through a sanitiser), and never run as code, SQL or a shell command. List what you changed.

Why vibe-coded apps leak

Lovable, Bolt, v0 and Cursor build what you describe, and you test it with your own account. Everything looks right, because you only ever see your own data. The leaks show up when a second person signs up.

Many of these apps run on Supabase, where the browser talks to the database directly with a key that is public by design. That is safe only while row-level security is on for every table, with rules that check who is asking. Leave one table out, or write a rule that lets everyone through, and anyone who opens the browser's network tab can read it.

What builders found in their own apps this year

“My supabase tables had row level security off. In plain terms: anyone who opened the network tab in their browser could read every row in my database.”

“a normal logged in user could still see other people's stuff even with RLS “enabled.””

“i swapped the id in the url for a teammate's and it just handed me their stuff with no check at all”

“A few days ago I logged into ElevenLabs and found my entire credit balance gone. 121,000 credits burned in a single day”

The first three are about who can read data, which is why that group comes first in the checklist. The last one started with a key committed to git in a local settings file.

How to use this checklist

Tick what your app already does. For anything left, open the item and copy its prompt into your coding agent, or copy all the prompts for unticked items at once. Ask the agent to work through them one at a time and to tell you what it changed. Then check the important ones yourself: sign in as a second user and try to open the first user's data, and search your published site's JavaScript for your key's first few characters.

What to do first

  1. Any key that has leaked. If a secret key has ever been in your browser code, a public repo, a screenshot or a chat, revoke it at the provider and make a new one before anything else. Deleting it from the code doesn't stop whoever already copied it.
  2. Who can read your data. Row-level security on every table, and the test as a second user. The first three posts above all came from here.
  3. Every request checks who's asking. Owner checks on every route that takes an id.
  4. Abuse and surprise bills, before you open sign-ups or hand out free credits. The AI launch checklist covers the rest of the AI cost side.
  5. AI features, once an AI step can call tools, send messages or read other people's data.

Run your builder's own scan too

The main builders now check some of this for you. Their scans and this list overlap, and neither catches everything:

  • Lovable runs a quick security scan every time you publish, and a deeper one on request.
  • Bolt has a security check, with a database check on every plan.
  • Base44 has a security scan for permissions, exposed keys and dependencies.

A scan reads your code and settings. The second-user test shows what a real person can actually reach, so do both. If you handle payments, health data or other businesses' data, pay for a professional review before launch.

Questions

Is the Supabase anon key safe in the browser?

Yes, it's meant to be public, and so is the newer publishable key. But it's only safe while row-level security is on for every table the browser can reach, with policies that check the user. The service_role and secret keys are different: they skip row-level security and must never leave the server.

Does this replace a security review?

No. It covers the mistakes that most often hurt vibe-coded apps. Run your builder's own security scan too, and if you handle payments, health data or other businesses' data, pay for a professional review.

I built with Lovable, Bolt, Replit, v0 or Cursor. Does this apply?

Yes. The checks are about what your app allows, not the tool that wrote it. Each item has a prompt you can paste into that tool's agent.

What if a key has already leaked?

Revoke it at the provider first, then replace it everywhere it's set. Deleting it from your code or your git history doesn't stop anyone who already copied it.

Is my progress saved?

Your ticks are saved in this browser only.