Can You Build a Website with AI? The Pros, Cons and Risks
Oblix Digital
Web design & SEO

Vibe coding means describing what you want to an AI tool such as Claude, accepting whatever code it generates, and shipping it without really reading or understanding what it wrote. The term started out as a slightly tongue-in-cheek description, but it's stuck because it describes something a lot of people are genuinely doing. I use AI to build every site I make, so I'm not here to dismiss it. I just think it's worth being clear about where it works and where it quietly goes wrong.
Where vibe coding genuinely shines
For prototypes, it's brilliant. If you want to see whether an idea has legs, test a layout with a few customers, or build a quick internal tool that only you will use, getting something working in an afternoon is a real advantage. If it breaks, nothing important breaks with it, and you've learned something cheaply.
The problem starts when real customers are involved
The risk isn't that AI always writes bad code. It's that nobody knows whether it did. A contact form can look perfect and still be wide open to spam bots. A booking system can work fine in testing and store customer details somewhere it really shouldn't. These issues don't announce themselves, they sit there until someone finds them, and that someone isn't always friendly.
Leaked API keys and open databases
The most expensive mistakes are the ones you can't see on the page. A website that takes payments, sends emails or saves bookings uses secret keys to talk to those services. If a key ends up in the code that loads in a visitor's browser, anyone can open the page source and copy it. Bots scan public code for exactly this, and a stolen key can be used to send spam from your account, run up usage bills or reach your data.
This happens at scale. GitGuardian counted 23.8 million secrets, such as passwords and API keys, leaked on public GitHub in 2024, up 25% on the year before.
The bills can be large. In March 2026 The Register reported on a developer charged $82,314 for Google's Gemini AI services after an unknown thief got hold of their API key. The report did not say how the key was taken.
The second risk is the database. Many AI app builders save your data in a hosted database that relies on access rules to keep it private. If those rules are missing, anyone who knows where to look can read, and sometimes change, every record: names, email addresses, phone numbers and booking details. Supabase, a database service many AI builders use, tells developers to switch on its access rules, called Row Level Security, for every table the website can reach, and warns that a table without them can be read and changed by anyone who has been given access to it. In 2025 a critical vulnerability of this kind (CVE-2025-48757, rated 9.3 out of 10) was published for sites generated with the AI builder Lovable, because missing access rules let anyone read or write their database tables. Lovable disputes the report, saying each customer is responsible for protecting the data in their own app.
For a UK business, a leak of customer details is a personal data breach under UK GDPR. If it is likely to put people at risk, you must report it to the ICO as soon as possible and, where feasible, within 72 hours. If the risk to people is high, you must tell the customers affected too. The ICO can fine up to £17.5 million or 4% of worldwide annual turnover, whichever is higher. Fines that size are rare, but the time spent investigating and fixing a breach, and the trust you lose with customers, can be costly enough on their own.
There is also a yearly fee to keep in mind. Organisations that use personal information, sole traders included, need to pay the ICO a data protection fee unless they are exempt, and a website that collects enquiries is handling personal information. The fixed penalty for not paying is £400 for micro organisations, £600 for small and medium ones and £4,000 for large ones.
What to check before going live
No secret keys in anything the browser loads. They belong on the server, in environment variables.
Database access rules switched on for every table, and tested while logged out.
Each key limited to what it needs, with spending caps or usage alerts on paid services.
A plan for what you would do, and who you would call, if a key leaked.
Nobody understands it when it breaks
Every website eventually needs something fixed. When the code was generated without anyone understanding it, that fix turns into guesswork. You ask the AI to repair one thing, it changes three others along the way, and you've no reliable way of knowing what else has moved. The original build was quick, but each change after that makes the whole thing a little more fragile.
Six months later is the real test
A site isn't finished at launch. Your prices change, you add a service, a dependency needs updating, or a browser update changes how something displays. Vibe-coded projects rarely come with any plan for this (there is rarely any documentation, clear structure or testing), so the cost of the original shortcut tends to show up later, usually at the least convenient moment.
Where I'd actually draw the line
Fine to vibe code: prototypes, mock-ups, personal projects and internal tools with no sensitive data
Needs someone reviewing the output: your actual business website, and anything with forms, logins, payments, API keys or customer data
Worth asking before you commit: who will maintain it, and whether they understand what's in it
The difference between vibe coding and using AI properly isn't the tool. It's whether someone reads, tests and understands what's going live, and that's the part I'd never skip.
If you've built something quickly with AI and aren't sure whether it's safe to put in front of customers, get in touch and I'll take an honest look at what's there.
A quick note: this is my own opinion, based on what I've seen work in practice, not formal legal, financial or professional advice. Your situation may differ, so do your own research before making a decision. These views are mine alone and don't reflect any employer or organisation I'm affiliated with.
Sources
- GitGuardian: The State of Secrets Sprawl 2025
- The Register: Dev stunned by $82K Gemini bill after unknown API key thief goes to town (3 March 2026)
- Supabase docs: Row Level Security
- NIST National Vulnerability Database: CVE-2025-48757
- ICO: Report a personal data breach
- ICO: The maximum amount of a fine under UK GDPR
- ICO: Data protection fee
- ICO: Fixed penalties for failure to pay the data protection fee
- Google Search Central: Spam policies
- OWASP Top 10: Web Application Security Risks
- NCSC: Small Business Guide
Related reading
Frequently asked questions
What does vibe coding actually mean?
Vibe coding means describing what you want to an AI tool such as Claude, accepting the code it generates, and shipping it with little or no review of what was actually produced. The focus is on whether it appears to work, not on understanding how it works.
Is vibe coding safe for a business website?
Not on its own. It's fine for a prototype, but a live site that handles enquiries, bookings, payments or customer data should have someone who understands the code reviewing its security, accessibility and performance before launch.
What happens if an AI-built website leaks customer data?
A leak of customer details is a personal data breach under UK GDPR. If it is likely to put people at risk, you must report it to the ICO within 72 hours and may need to tell the people affected. The ICO can fine up to £17.5 million or 4% of worldwide annual turnover, whichever is higher.
Can a vibe-coded website be fixed and made production-ready?
Usually, yes. A developer can review the code, close security gaps and tidy up the structure so it can be maintained. Depending on how it was put together, rebuilding the key parts properly is sometimes quicker than patching everything.
Is building a website with AI the same thing as vibe coding?
No. Vibe coding is shipping AI output without understanding it. Using AI deliberately means a developer directs it, reviews each change and tests the result before it goes live. The tools can be identical, the outcome is very different.


