In Britain, very few buildings just fall down. Even in stormy weather, or when we have the odd tornado or mild earthquake. With only a few exceptions - gas explosions, mostly - there’s not a lot that will cause a building in Britain to simply collapse.
The same cannot be said for buildings in some countries.
The difference is one of engineering. We don’t just stack up whatever building material comes to hand and hope. There’s a process, and it involves architects, structural engineers, materials scientists, a strict code of building control, and independent surveyors and building inspectors. Every one of them is there because somebody, somewhere, once learned the hard way what happens without them.
Even the gas explosions taught us something. In May 1968, a gas explosion on the eighteenth floor of Ronan Point, a brand-new tower block in Canning Town, blew out a load-bearing wall, and an entire corner of the building fell down like a stack of playing cards. Four people died. Britain’s response wasn’t to shrug. It was to change the building regulations, so that one failure could never again bring down everything stacked on top of it. That’s what engineering is: a discipline that remembers its disasters and refuses to repeat them.
Now picture a city where anybody can put up a building. No plans, no permission, no inspection. Just describe what you’d like to a very obliging machine, and it pours the concrete overnight. It looks magnificent. The lights come on. Nobody checks the foundations, because nobody knows there’s anything to check.
Welcome to vibe coding.
Software without the pesky engineering
Vibe coding is software building without all of the pesky engineering. You describe what you want to an AI, it writes the code, the code runs, and you ship it. Job done.
Except it isn’t. I’ve been building software for a very long time, and I have never seen an application created this way that will stand up to any real-world challenge at all. Not one. It works in the demo. It works for the person who made it, with their data, on a good day. Then it meets real users, real data and real attackers, and it falls over - or worse, it doesn’t fall over, and just silently, imperceptibly… leaks.
This isn’t a hunch. When Escape.tech scanned around 5,600 publicly deployed vibe-coded applications, they found more than 2,000 high-impact vulnerabilities, over 400 exposed secrets - API keys, access tokens and the like - and 175 instances of personal data, from medical records to bank details, sitting out in the open. Researchers at Carnegie Mellon put a leading AI coding agent through a batch of real-world programming tasks: its code worked 61% of the time, but it was secure only 10.5% of the time. And the Tea app, a women’s safety app, managed to spill tens of thousands of photos and ID documents, and over a million private messages, onto the internet in 2025.
It runs. That’s not the same as it being safe. See the problem?
Brilliant, and still cutting corners
Let’s be clear about one thing, because people get it wrong in both directions. The models have got really, really good. The best of them can now do genuinely incredible work - reason through a gnarly bug, untangle a messy module, write a decent test suite, all in minutes. I work alongside them every day, and some days they take my breath away.
But they still take shortcuts. They still miss things. Ask one to check a thousand lines of code and it’ll often read the first hundred carefully, skim the rest, and tell you with total confidence that it’s all fine. They summarise rather than look at the detail. They’ll write a test that passes because it doesn’t really test anything. They’ll fix the symptom you described and leave the cause untouched. And all of it comes wrapped in perfectly fluent, perfectly reassuring prose.
On its own, that’s manageable. A good engineer catches it, the same way a good site foreman catches the apprentice who’s skimped on the wall ties. The danger is what happens when there’s no foreman. Because AI development isn’t one pass - it’s a loop. The agent builds on what it built yesterday, which was built on what it built the day before. Every shortcut becomes the foundation for the next iteration, and the next iteration doesn’t look underneath. So any laziness in the surrounding engineering - no review, no meaningful tests, no architecture holding the thing in shape - doesn’t just let those problems survive. It amplifies them, over and over again, iteration after iteration.
One weak joint, and everything stacked on top of it. Sound familiar?
We’ve been here before - sort of
There has always been shadow IT inside corporates. Little enclaves of functionality built in spreadsheets and in low-code tools like Microsoft Access. Users will always find a way, and when the people whose job it is to solve business problems don’t solve them, users go and build their own solutions.
A very long time ago, in a former life, I had the job of documenting and replacing huge, out-of-control, monolithic systems built in Lotus 1-2-3 by that rare breed, the citizen developer. Typically someone from management accounts with a penchant for DIY and an unhealthy disregard for documenting anything at all. It was miserable work. But here’s the thing: those systems were mostly harmless. Old shadow IT was hard to keep track of, but it was unconnected and isolated. A spreadsheet on a shared drive, used by a small team to lubricate the wheels and speed their work along. If it broke, it broke on one desk. If it was wrong, it was wrong in one department.
It was a garden shed. Badly built, probably, but if it fell down, it fell on the lawnmower.
This time it’s wired into the mains
Hand AI development tools to users, with no engineering rigour enforced, and really bad things can happen. Because vibe-coded applications aren’t garden sheds. They’re connected - to your other systems, to the internet, and to essential APIs.
Think about what an enthusiastic, well-meaning person in your sales or finance or operations team can build in an afternoon now. A little web app that pulls customer records out of the CRM. A tool that reads the shared mailbox and files invoices. An “agent” with a Microsoft 365 login and permission to read, write and send on someone’s behalf. A dashboard published on a free hosting account so the team can see it from home - which also means anyone can see it from anywhere.
To make those things work, the AI cheerfully asks for credentials, and the user cheerfully hands them over. Admin API keys end up hard-coded into code that runs in the browser. Database passwords end up in public repositories. Customer data ends up on a free-tier database in a jurisdiction nobody chose. And none of it has been looked at by anyone who knows what an injection attack is, what least privilege means, or why you don’t let a web form talk directly to the finance system.
The shed’s been wired into the mains. With bell wire. By someone who’s never heard of a fuse.
And now it’s company policy
If only it were just sheds. It’s worse than that.
Old shadow IT was, by definition, in the shadows. It lived in little enclaves because it knew it wasn’t really allowed, and that kept it small. This time, in some large corporates, it isn’t in the shadows at all. It’s been written into software development policy.
Managers with no software engineering experience, no awareness of what the discipline involves and no knowledge of what can go wrong are being handed Claude Code, Codex and the rest, and told they can now write their own applications. Not as an experiment. As the way things are done. “This is how we do software now,” one proudly proclaimed.
Remember that imaginary city, where anybody can put up a building overnight with no plans and no inspection? Some corporates have just voted to become it. This isn’t the odd cowboy builder working on the quiet. It’s the council abolishing building control and handing every resident a cement mixer.
And think about what that does to everything I’ve described so far. A sanctioned app doesn’t stay small - it’s encouraged to grow, to connect to the production systems, to be rolled out to whole departments, because the policy says that’s fine. Every one of those managers is now running their own loop of amplified shortcuts, with nobody qualified checking a single iteration of it. And the people with the authority to say “stop” are the very people doing the building. The engineers who raise an eyebrow get told they’re dinosaurs defending their jobs.
Nobody’s skimping on the wall ties in secret any more. They’ve been told wall ties are optional.
And you can’t see any of it
These applications are every bit as hard to detect and track as the old shadow IT. Arguably harder. They don’t live on a shared drive where someone might stumble across them. They live on personal hosting accounts, in browser extensions, in scheduled tasks on laptops, in automation platforms signed up for with a company email address and a personal credit card. They don’t appear in your asset register. They don’t appear in your security team’s threat model. They didn’t go through change control, because nobody thought of them as change.
And when the person who built one leaves the company? It keeps running. With their credentials. Indefinitely.
The sanctioned ones are no better. Everyone knows they exist; nobody knows what’s in them. Knowing there’s a building on the corner isn’t the same as knowing whether it’ll stay up.
So here’s an uncomfortable question. How many of these are running inside your organisation right now? Not roughly. Actually. If you can’t answer, you’re in the same position as a city council that’s just discovered half its streets were built overnight by people with no plans - and has no idea which half.
Banning it won’t work. Engineering it will.
The reflex response is to ban it. Lock down the AI tools, send a stern memo, job done. Alas, that’s never worked with shadow IT and it won’t work now. Users found a way round the IT department with Lotus 1-2-3 decades ago; they’ll find a way round it with an AI in 2026, only faster. The need doesn’t go away just because you’ve forbidden the solution. And where vibe coding is company policy, there’s nobody left to issue the ban - the people who’d sign the memo are the ones writing the apps.
Nor is the answer to throw the technology out. The models really are extraordinary - we use them ourselves, every day. The problem was never the machine. The problem is building without engineering. Britain didn’t stop building tower blocks after Ronan Point. It started building them properly.
That’s what’s needed here. Building control for software. And it needs doing now, before your organisation has its own Ronan Point moment - a breach, a regulatory fine, a headline with your company’s name in it.
Find it, fix it, and make it safe
This is exactly what we do at JFDI. We’re engineers. We’ve spent decades designing, building, rescuing and replacing business software, and we’ve spent the last few years working hands-on, daily, with the same AI tools your people are using - so we know precisely where they cut corners, because we’ve watched them try.
We find the bad stuff. A proper survey of what’s actually out there: the vibe-coded apps, the rogue agents, the exposed keys, the over-privileged connections, the data that’s wandered off somewhere it shouldn’t be. You can’t fix what you can’t see.
We fix it. Everything we find gets triaged by risk. Some of it gets secured. Some of it gets re-engineered into something robust and maintainable. Some of it gets switched off - gracefully, with a proper replacement, so the people who depended on it aren’t left stranded.
We fix the process. This is the bit that matters most. If your organisation has made AI development its policy, we help make engineering part of that policy too. We put the engineering back into AI development: architecture before code, secrets kept out of source, least-privilege access, automated testing and security scanning, independent checks that never take the model’s word for its own work, human review at the go/no-go gate, and a governed route for your people to get what they need built - safely. Your users keep their superpowers. You stop collecting time bombs.
We make it safe, and keep it safe. Independent inspection, not a one-off tidy-up. The surveyor and the building inspector, for software.
The good news is that none of this is new. We know how to build things that don’t fall down. We’ve known for a long time. We just have to bother to do it.
Every day you wait, somebody in your organisation builds another one. Get in touch with us at JFDI, and let’s find out what’s standing on your foundations before it falls down.
How many vibe-coded apps are running on your foundations?
We find the vibe-coded apps, rogue agents and exposed keys inside your organisation, fix or retire them by risk, and put real engineering back into how your people build with AI.

