Skip to content
ScanAnt

Vibe coding security scanner for AI-built apps

When an AI writes the app, the usual mistakes change shape: model keys pasted into the front end, an AI endpoint anyone can call on your bill, a paywall that exists only in the interface. ScanAnt looks for these AI-app patterns next to the standard web checks, whichever builder you used.

Free, no signup. We only read what your site already shows every visitor.

What we check

Each item is tagged with the scan that reports it. The quick scan is free and needs no signup. The others run only on sites you have proved you own.

  • AI provider called from the browser

    Quick scan

    A model API reached directly from front-end code, which means its key is visible to everyone.

  • System prompt in public code

    Quick scan

    Instructions for your AI feature shipped inside the page, which anyone can read and copy.

  • AI endpoint that looks open

    Quick scan

    From public code, an endpoint that calls a model without sign-in or limits, which spends your money.

  • Prices, roles and paywalls decided by the client

    Quick scan

    The browser sets the amount it pays, flags itself as admin or unlocks paid features with a local variable.

  • Anonymous requests to AI endpoints

    Deep scan

    GET and OPTIONS requests to found endpoints show whether authentication is enforced and whether CORS allows other websites to post to it. We do not send a model request.

  • Optional POST probe

    Deep scan

    Only with a separate opt-in, up to 3 POST requests to see whether an AI endpoint answers anyone. It can spend a small amount of your credit, and you are told before it runs.

  • Database and storage access

    Verified scan

    Tables, collections and files readable without sign-in on Supabase or Firebase backends.

  • Rate limits in your code

    Code scan

    In an uploaded project: AI endpoints without a per-user limit, limiters stored in one process's memory and public endpoints without abuse protection.

  • Quick scan: No signup. Reads only what every visitor receives.
  • Deep scan: After you prove ownership and tick the deep option.
  • Verified scan: After you prove you own the site. Read-only.
  • Code scan: You upload your project as a ZIP.

No scan can promise to find every problem. A clean result means no known problems were found on what the scan could see.

Common mistakes

  • Calling a model API from the browser

    The key and every prompt are exposed, and anyone can run up your bill.

  • Shipping the system prompt to the client

    It reveals your product logic and makes jailbreaks easier.

  • An AI route with no sign-in and no per-user limit

    A script can loop it until the monthly credit is gone.

  • Checking the plan in the interface only

    A user can edit local state and use paid features free.

  • Relying on the builder's default database rules

    Generated policies are often too open, and nobody reads them.

How to fix it

  1. Route every model call through your server and keep the key there.

  2. Require sign-in on the AI endpoint, add a per-user and per-IP limit and set a spending cap at the provider.

  3. Keep the system prompt on the server and validate the user input that goes into it.

  4. Move price, role and entitlement checks to the server or the database policy.

  5. Rescan after each builder prompt that touches auth or data.

Questions

Does the scanner call my AI endpoint?

Not with a model request unless you separately opt in. By default it reads public code and sends GET or OPTIONS requests only, so it spends none of your credits.

Which AI builders does it support?

It is built for apps from Lovable, Bolt, Replit and Cursor, but any web app that ships public JavaScript or a reachable server can be scanned.

Can it prove that my AI feature is secure?

No. It reports likely and confirmed open paths and says when protection could not be confirmed without spending money. Treat it as a way to find the obvious holes fast.

What does "likely" mean in a finding?

From public code alone we can see the pattern, such as a call with no auth header, but not the server's answer. Findings are worded accordingly, and a deep scan can confirm them.

How often should I rescan?

After every significant change by the AI. Daily monitoring is planned but is not available yet.

Run the scan now

Free for any public site, no signup. Need the verified, deep or code scan? Choose the level on the scan page.

Choose a scan level →