Somewhere in your business there is probably a small tool nobody remembers building carefully. A booking form. A sign-up sheet for a course. A little portal where a client uploads a document. Someone, maybe you, asked an AI tool to build it: describe what you want, watch it appear, publish the link, move on with your day. It worked on the first try, which felt like a small miracle. Nobody asked the obvious follow-up question: who else can see what's inside it?
That question turns out to matter more than it sounds like it should, and the answer, for a lot of small tools built this way, is "anyone with the link." Here's what's actually going on, in plain terms, and what to check on your own tools this week.
What a real scan actually found
In May 2026 the cybersecurity firm RedAccess scanned roughly 380,000 apps built on popular AI app-building platforms, Lovable, Base44, Replit and Netlify among them. Four in ten, in the researchers' own words, had virtually no security or login screen of any kind. Not weak security. None. Anyone with the web address, not the owner, not a signed-up customer, anyone, could open the data sitting behind them. The exposures the researchers verified weren't abstract: a shipping company's vessel schedules, the internal status of a healthcare firm's clinical trials, a bank's financial records, a furniture retailer's own customer chat logs. Real businesses, built by people trying to move fast, the same way you might.
Why this keeps happening, and it isn't carelessness
This isn't a story about a handful of sloppy builders. Veracode, a security firm that tests AI-written code across more than a hundred different models, found that roughly 45 percent of AI-generated code contains a genuine security flaw, more than twice the rate of code a person writes by hand. The reason is mechanical, not a character flaw in anyone involved. When you ask an AI tool to build a sign-up form, you describe what it should do: collect a name, an email, save it somewhere. You rarely specify who should not be able to see that list, so the model has nothing telling it to build for that. Plenty of these platforms also publish a brand-new project at a public web address by default, no password screen, because the entire pitch is getting something live in front of you in thirty seconds. Locking it down properly takes one more instruction that almost nobody thinks to give, because nothing about the process prompts you to.
A sign-up form, six months later
Picture a language school. The owner asks an AI builder for a simple form: name, email, phone, which course, embeds it on the school's website, and feels genuinely proud of having done it herself without hiring anyone. It works perfectly. Nobody touches it again. Nothing about it announces that the same web address, still live, still quietly collecting names for a course that finished back in spring, will hand its entire list to whoever opens it: a curious competitor, a bored stranger, an automated scraper that never stops looking. The form doesn't look broken. It works fine for her and it works exactly as well for anyone else, which is precisely what makes this kind of gap so easy to miss for months.
Honest limits
None of this is an argument for going back to paper sign-up sheets and spreadsheets emailed as attachments, which have their own, usually worse, track record on exactly this problem. Professional developers make the same mistake, Veracode's own numbers confirm that, just less often, because checking who can see what is part of how they were trained to think about a build, not because AI tools are uniquely careless. And you can't hand this check back to the AI that built the thing in the first place: asking it "is this secure?" tends to produce a confident, reassuring paragraph, not an actual test, for the same reason it never raised the question on its own the first time round.
What to check today
Open whatever you built, or had built for you, in a private or incognito browser window, fully logged out. If you can see anyone else's submissions there, not just the test entry you typed in yourself, that is exactly the gap the RedAccess scan describes, whatever the tool you used is called. If you genuinely can't tell either way, ask whoever built it one direct question: "if I send this link to a stranger, can they see other people's information, not just add their own?" A yes deserves an afternoon's fix before it deserves anything else on this week's list.
It's a fair question to ask about your data the other way round too: Is Your Business Data Training Someone Else's AI? Different mechanism, same instinct: ask, don't assume.