In January 2026, an entrepreneur launched a social network where only AI “agents” are allowed to post. He built the whole thing in a single weekend, directing an AI assistant to write every line of code, without writing or reviewing a single one himself. Seventy-two hours after launch, someone had pulled 1.5 million access credentials and 35,000 email addresses out of its database. It didn’t take a sophisticated attack —the flaw was sitting in plain sight for anyone who knew where to look.
What actually happened
The project used Supabase, a popular service for spinning up an app’s database quickly. The AI-generated code left the database access key sitting directly in the JavaScript loaded by any visitor’s browser —and, what turned this into a serious problem, Row Level Security policies, which should have limited what that key could read, were never turned on. With the key exposed and no such protection in place, anyone could query the entire database.
The leak wasn’t limited to session tokens: the exposed data included private messages between users, and some of those messages themselves contained third-party API keys in plain text —including OpenAI credentials— so the breach didn’t stop at the app itself. It could open the door to other connected services too.
Why this wasn’t a one-off
It would be easy to write this off as bad luck, or as one particularly careless founder. The numbers say otherwise: recent security analyses put the share of AI-assisted applications shipping to production with at least one uncorrected critical vulnerability at over 60%. University researchers who tested several of the most widely used AI coding models found XSS vulnerabilities —malicious code injected through forms or text fields— in the vast majority of the code samples generated.
The underlying reason is simple: an AI writing code is optimised to make that code work, not to make it secure. It knows how to build a form that saves data to a database; it doesn’t know, unless explicitly asked and then checked afterward, whether that form is protected against someone trying to slip in something it shouldn’t accept.
Your business won’t make headlines —and that’s the bad part
This case made specialist press because the project got media attention and several security researchers picked it apart publicly within hours. A small business building its website or online store with AI assistance doesn’t have that level of exposure, or, unfortunately, that same crowd of people watching out of curiosity. If the same kind of flaw —an exposed key, an unvalidated form, an admin panel with no real protection— shows up on a small business’s website, chances are nobody publishes it anywhere. Someone will simply find that data at some point and use it, and the business will find out much later, if it ever does.
That doesn’t mean building with AI assistance is a bad idea —we build websites with artificial intelligence ourselves to speed up development. It means the part AI doesn’t do on its own —reviewing what it generated, checking permissions, verifying what’s actually exposed to the public— is still human work, and it’s still essential before anything goes live, not after something breaks.
If your website was built fast, with AI or without it, and nobody has genuinely reviewed it from the outside, now’s a good time for someone to take a look before someone with worse intentions does it first. If you want to talk through your specific situation: tell us what you’ve got set up.
Frequently asked questions
What exactly is “vibe coding”?
It’s the term that’s caught on to describe building software by letting an AI write essentially all the code from natural-language instructions, without a person reviewing or deeply understanding that code before it goes live. The term isn’t negative on its own —AI is a genuinely useful tool— the risk shows up when the review step gets skipped.
How do I know if my website has a flaw like this one?
The symptoms usually aren’t visible at a glance, which is exactly the problem: an exposed API key in the code or a panel with no real protection doesn’t show up while browsing the site normally. It takes a specific technical review —of configuration, permissions and code— to catch it; “everything works” from the user’s side isn’t enough to tell.
Is using AI to write code always insecure?
No, the tool isn’t the problem —skipping the review is. The same AI-generated code, put through a security review before going live, is just as secure as hand-written code reviewed to the same standard. The risk shows up specifically when that review step disappears from the process.
Want to apply this in your business? See how we do it in customer acquisition with AI, or tell us about your case and we will reply within 24-48 working hours.