Built your app with AI? Do the logged-out test before customers sign up

An app can look finished and still hand out customer rows to anyone who opens it. Here's the short test I run, what a German security firm's study reported on 2 November, three checks that matter, and when to bring in a developer.

Side view of a silver laptop on a light wooden desk in warm natural light, ready for a logged-out security check

I keep meeting the same product. Nice screens. Sign-up form. A dashboard that loads customer names in under a second. Then I open it in a private window, still logged out, and the same names arrive anyway. The buttons look locked. The data is not.

The short answer

If your app stores anything you would not put on a public homepage (emails, bookings, invoices, chat, files), treat "it looks done" as half the job. The other half is making sure a stranger who never creates an account cannot read those rows. That is not a rare edge case this month; it is the pattern behind a study that landed in the German press on 2 November 2026, and the part that should worry you is how ordinary the cause looks once someone walks through it without logging in.

The logged-out test anyone can do

You do not need a security certificate for this. You need twenty minutes and a browser that is not already signed into the app.

  1. Open a private or guest window. Do not log in. Do not paste a session cookie. Go to the public URL customers would use.
  2. Click through the marketing pages and any "try it" screens the way a curious stranger would. Fill nothing that creates an account unless the product forces it, and if it forces it, stop and ask why a demo needs a real inbox.
  3. Open the browser developer tools, Network tab. Reload. Watch for requests that return lists of people, messages, orders or files while you are still anonymous.
  4. On your phone, repeat the visit in a browser that has never seen the app. Same check.

If you see rows that belong to other people, or fields you would email a lawyer about, close the tab and treat it as a live problem. Do not keep clicking around "to see how far it goes". Do not publish a how-to. Just note what loaded while logged out, and who can fix the access rules.

I am not asking you to attack anyone else's product, and I am not giving you a recipe for poking at strangers' databases; this is your app, or one you are about to put paying customers into, and the only goal is to see what it already answers while you are anonymous.

Why this happens: the screens are built, the locks aren't

AI app builders are good at the visible part. Forms, lists, pretty empty states. The database behind them ships with a public key in the page on purpose, so the browser can talk to the backend. That design is fine only when every table has clear rules about who may read or write which rows, and the moment those rules are missing the public key in the page stops being a convenience and starts being a open front door.

When those rules are missing or too loose, the page can load "public" data for the demo and, with the same key, answer further read requests for tables nobody meant to expose. No break-in required. No exotic exploit. No stolen password. The app simply answers because the locks were never written down in the first place.

On 2 November 2026, Handelsblatt reported work by the Hamburg firm Crissy Field on projects from one popular AI app builder. The firm's own write-up says it analysed 159,041 projects found in Common Crawl between December 2025 and September 2026. Of those, 34,545 used one widely paired hosted database. 24,664 sites sent database records on first load without a login. That alone is not always a leak: some data is meant to be public. With further read-only requests using the same key the page already hands the browser, the firm found records containing email addresses in 10,560 tables across 7,403 of those 24,664 sites (about 30%).

Read their caveats with the numbers. They say the 30% figure is not a confirmed leak rate, and the Common Crawl sample is not representative of every app on that builder. Emails can be public contact addresses or test junk; personal rows without an email field were missed. They did not publish access recipes. Cause, in their words: missing or loose access rules in the database, not a hack of the builder itself.

So the headline is not "AI apps are doomed". Simple point. Finished-looking apps still need locks, and a lot of vibe-coded ones shipped without them.

Three checks before launch

1. Access rules on every table that holds real people. For each table that stores users, bookings, messages, files or payments, ask: what can an anonymous visitor read? What can a signed-in user read that belongs to someone else? The answer you want in writing is "nothing that is not meant to be public", with the rule names, not a shrug. Hosted databases used by these builders call this row-level security or policies, and if a table sits in an exposed schema with no rule at all, you should treat that as a stop-ship until someone writes the rule and proves an anonymous visitor cannot read another person's rows.

2. No secret keys in the browser. The public anon key that the page ships is expected. API keys for mail, payments, AI models, admin dashboards are not. Search the shipped front-end for strings that look like secrets. If the app can charge a card or call a paid API without going through your own server, pause right there, because a key that lives in the browser is a key that any visitor can copy out of the Network tab in under a minute.

3. A second pair of eyes. Someone who did not build the screens should try the logged-out test and read the access rules. That can be a developer you already trust, or us. It is cheaper than explaining a customer dump to your insurer. For a wider website checklist after the October PHP alert, see the website security check. If you are still choosing between a builder, a subscription product and a custom build, put this test in the decision, not after launch. Mobile and desktop apps have the same rule: a pretty client does not replace server-side checks (mobile app costs, and our software page).

If you find a problem

Take the app off public URLs if real customer data is already in it, or freeze new signups. Rotate any keys that lived in the browser. Tighten the access rules so anonymous reads return only what a brochure would show. Tell affected people if personal data was exposed. The legal bit depends on where you and they sit, so get proper advice; I am not your lawyer. Then re-run the logged-out test until the Network tab is boring.

Do not announce a "breach how-to" on social media. Do not paste customer rows into a bug report without stripping them. Fix first. Write the post-mortem second.

When to bring in a developer

Stay with the builder if the product is still a demo, the data is fake, and you are willing to learn the access-rule UI before anyone real signs up.

Bring in a developer when the app already has real customers, when you cannot explain who can read which row, when payments or identity documents are in scope, or when the builder's defaults keep fighting you. A short review of rules, keys and the logged-out path is ordinary software work. A full rebuild is only needed when the data model itself is a mess, and that is a separate conversation.

If you want that second pair of eyes, email the app URL (and a test login if the interesting screens sit behind one) to contact@magnuszgr.com. I will tell you what I would check first. I will not invent a fear score, and I will not ask for production admin access on a first mail.

The screens can look done on Monday. Locks matter more. Make sure they are done before Tuesday's signups.

Read next