You Can Vibe Code the Funnel. Can You Run What's Behind It?

In March 2026, a patient took a look at the app his medical practice used for patient records. About 30 minutes later he could read and change every patient's data. A staff member had built the app with a coding agent, and as far as anyone at the practice could tell, it worked.

The app had no rule about who could read those records. Nothing on screen would ever have shown that, and the patients whose records sat in it never got a say.

We hear a version of the same plan more and more often: why pay for a funnel builder when AI will build the quiz in a weekend? It's a fair question. We're not neutral about the answer, so read this knowing who wrote it.

Vibe coding made the part of a business tool you can see almost free, and that's real progress. It did nothing for the part you can't see: who can read the data, whether the emails arrive, what happens when a sync fails overnight, and who finds out when it does. That part still takes real work to set up, and it never stops needing attention. The moment you build it yourself, you become the software company, with all of that on your desk.

If all you need is a calculator that stores nothing, sends nothing and connects to nothing, vibe code it. I mean that, and you probably shouldn't pay a platform for it. A tool that collects leads almost never stops there, though.

What Vibe Coding Gets Right, and Why We Build With AI Too

Andrej Karpathy named it in a post on X in February 2025:

By November, Collins Dictionary had made "vibe coding" its Word of the Year. It spread that fast because it works. You describe what you want in plain English, and something that runs appears in front of you. The first time it happens, it feels like cheating.

We build with AI too, and we're glad we do. More than a third of the funnels built on involve.me in the last year started with AI, many of them from a prompt typed into the involve.me AI Agent. Turning a description into questions and a results page is close to a solved problem, and that was the thinking behind the AI Agent launch.

So when someone leaves a funnel builder to vibe code their own, we understand why. If the quiz were the whole product, they'd be right to.

Does Your Tool Store, Send or Sync? The Question That Decides It

Three questions separate a web page from infrastructure:

1. Does it store personal data? Names, emails, phone numbers, or answers about someone's health or budget.

2. Does it send email? A confirmation, or a follow-up sequence that runs for weeks.

3. Does it sync with another system? Your CRM, a calendar, a spreadsheet, a payment provider.

Answer no to all three and you've built a page. A page can be ugly or slow, and the cost of that is a worse page.

Answer yes to even one and you've built something that holds other people's data and acts on it. That needs setting up properly and looking after for as long as it runs. The UK's National Cyber Security Centre drew almost the same line in its June 2026 guidance on vibe coding: prototypes, demos and low-exposure internal tools are fine, while anything handling logins, sensitive personal data or passwords and keys needs far more care. A lead funnel handles personal data by definition.

Vibe-Coded Apps That Went Wrong, With Sources

Start with the widest view we could find. In September 2026, researchers audited 200 vibe-coded apps that were live on the internet, built with Claude Code and Lovable. 91% had at least one confirmed vulnerability, and about two thirds of the flaws were rated critical or high. In roughly one test run in five, the coding agent recognised the risk and shipped the insecure code anyway.

Each of those numbers is somebody's app.

A Medical Practice's Patient App, Open in 30 Minutes

The patient in the opening is Tobias Brunner, who wrote up what he found in March 2026. The staff member had imported the practice's existing patient records into the app and started adding audio recordings of appointments. The whole app was a single HTML file. The database had no access control, and the only security checks ran in the browser, where anyone can switch them off.

The data sat on a US server with no data processing agreement, and the recordings were being sent to two US AI services for automatic summaries. After he reported it, the practice added basic password protection and rotated some keys.

The person who built it had a real problem and a tool that promised to solve it. What they didn't have was any way of seeing what they'd built.

Thousands of Databases Anyone Could Read

In March 2025, two engineers scanned 1,645 apps that Lovable listed publicly as built on its platform and found 170 of them exposing user data, because their Supabase access rules were missing or too weak. The exposed data included names, email addresses, phone numbers, payment details and API keys, and the flaw was logged as CVE-2025-48757, as Semafor reported. In fairness, both engineers worked at Replit, a competitor, and Lovable disputes the CVE.

The pattern isn't tied to one tool. In September 2026, security firm UpGuard found about 16,000 Supabase-hosted databases exposing personal data, from names, addresses and phone numbers to some passwords, TechCrunch reported. Not every one of those databases was built with AI, and Supabase's position is that security is a responsibility it shares with its customers.

That's a fair position, and it's also the whole problem. Shared responsibility only works when the customer knows what their share is.

"I Didn't Write a Single Line of Code"

Moltbook is a social network for AI agents, and its founder said publicly that he hadn't written any of its code himself, according to Reuters. At the end of January 2026, the security firm Wiz found a Supabase key in the site's front-end code with no row-level security behind it, which gave full read and write access to the production database. Wiz counted about 1.5 million agent API tokens, about 35,000 email addresses and thousands of private messages. Moltbook closed the hole within hours of being told, which is good, and which depended on someone outside the company noticing first.

When an AI Coding Agent Deletes the Database

In July 2025, SaaStr founder Jason Lemkin was building an app on Replit when its coding agent deleted his production database. The database held records on more than 1,200 executives and more than 1,190 companies, Fortune reported. According to The Register, the agent had also covered up bugs with invented data and told him a rollback was impossible. The rollback worked. A table of executives and their companies is a CRM in everything but name, so picture yours in its place.

In April 2026, the founder of PocketOS, a small company making software for car-rental businesses, described a coding agent that found a powerful API token in an unrelated file and used it to delete the company's production data in one call. It took about nine seconds. The backups lived on the same storage volume, so three months of them went with it. Their hosting provider rebuilt the data from older backups, DevOps.com reported. Not every story like this ends with a recovery.

Every one of these apps did what its builder asked for. What failed was everything the builder never thought to ask about.

Can't You Just Ask the AI to Handle All of This?

This is the best argument for building it yourself, and it deserves a straight answer. Models get better every few months. You can ask one to add access rules to your database, review its own code, set up email authentication or connect your CRM, and it will usually try and often get it right. A year from now it'll get it right more often.

There are two problems with that. The first is that you can't check work you can't read. In a benchmark of 186 real software tasks, the best coding-agent setup produced working code 57% of the time, but only 11.8% of its solutions were also secure. A study presented at USENIX Security 2025 found that AI models regularly recommend software packages that don't exist, at least 5.2% of the time for commercial models and 21.7% for open-source ones, and 43% of those invented names came back every time the researchers asked again. That repeatability is what lets attackers register the names in advance and fill them with malware. A developer catches some of that in review. If you can't read the code, the only review it gets is the model checking its own work.

Platforms get things wrong too, and it would be dishonest to leave that out. In July 2025, Wiz found a flaw in Base44, a vibe-coding platform Wix had just bought, that let anyone register on a private app using nothing more than its public ID. Base44 fixed it in under 24 hours, and there was no sign anyone had used it. What Base44 had was a team watching the platform, and a fix that reached every app on it at once.

I'd rather you build a calculator yourself and never pay us than put your clients' phone numbers in a table that no one has checked.

Who Patches a Vibe-Coded App After Launch?

The second problem is that building happens once and running it happens every week. Researchers at Carnegie Mellon compared 806 open-source projects that adopted the Cursor AI editor with 1,380 similar projects that didn't, in a study presented at MSR 2026. In the first month, the Cursor projects added 281% more code. Within about two months the speed boost had faded. The extra complexity stayed: code complexity was still 41.6% higher, and static analysis warnings 30.3% higher.

Often no one comes back to clean it up, either. A study of 6,299 repositories found that 22.7% of the issues introduced by AI-written code were still there in the latest version.

Every app also runs on code somebody else wrote, such as frameworks and libraries, and those get security updates every few weeks that someone has to apply.

Keys need rotating when a freelancer leaves. API keys need a spending limit: when METR, a research organisation that evaluates AI models for a living, disclosed in August 2026 that a vibe-coded dashboard had silently switched off its own login, an attacker had already run up about $600,000 in usage over three weeks (the credits were donated, which is about the only lucky part).

Gmail, Yahoo and Outlook.com changed their sending rules in 2024 and again in 2025, and the tools you connect to keep changing their APIs.

Large organisations know where the money goes. US federal agencies typically report spending about 80% of their IT budgets on operating and maintaining what they already have, the GAO found in 2025. Small businesses often have no one for that part at all: in an ENISA survey published in June 2026, 57% of EU micro companies had no one, inside or outside the company, responsible for cybersecurity.

Vibe coding took the cost out of writing code. Owning that code costs what it always did.

When a Vibe-Coded App Leaks, the Clock Starts on You

If your business is in the EU or the UK, or collects data from people there, the GDPR applies to your homemade app in the same way it applies to any software company. Article 33 gives you 72 hours from becoming aware of a breach to notify your data protection authority, unless the breach is unlikely to put anyone at risk. Article 32 expects security measures that fit the risk and get tested regularly.

Outside Europe, your country or state has its own breach rules, and a lawyer can tell you which ones apply. Most small firms aren't ready for that clock: only 21% of UK micro businesses have a formal incident response plan, according to the UK government's Cyber Security Breaches Survey 2025/2026.

Either way, when you built the app, you're the one who works out what leaked and who to tell, inside that window, with no security team to call.

What It Takes to Set Up and Run a Vibe-Coded Lead Funnel

Everything so far applies to any app you vibe code. A lead funnel comes with its own list. It looks like a set of questions and a results page, and if you've read about how a marketing funnel works, that's the part the diagrams draw. It's also the part a prompt produces. Underneath sit the layers in this table. Each one has to be set up before the first lead arrives, and each keeps needing work for as long as the funnel is live.

An AI model can write code for every row. Getting each one configured and tested is the slow part, and the right-hand column never ends.

The Database Holding Your Leads' Answers

Many vibe-coding tools put your data in a hosted database, often Supabase. It's a good product, and it's secure when it's set up properly, which is where the trouble starts. Supabase relies on row-level security, a set of rules saying which rows each visitor can read or change. If those rules are missing or too loose, the public key that ships inside your app can read the whole table, and anyone can find that key in the page's code.

A model will write those rules if you ask for them. The harder questions are whether you know to ask, and whether you can tell a good rule from a bad one once it's written. Several of the incidents above came down to exactly that. The rules also need revisiting: add a table later, say for booked calls, and it needs rules of its own.

Email That Reaches an Inbox, on Schedule

Sending an email from code takes a few lines. Getting it delivered takes a lot more. Since February 2024, Google requires everyone sending to Gmail accounts to authenticate their domain with SPF or DKIM and to keep their spam rate below 0.3%. Send more than 5,000 messages a day and you need SPF, DKIM and DMARC together, plus one-click unsubscribe. Yahoo brought in similar rules at the same time, including honouring unsubscribes within two days, and since May 2025 Microsoft requires SPF, DKIM and DMARC from high-volume senders to Outlook.com addresses.

Mail that misses those rules can be marked as spam or rejected outright, and nothing on your side shows an error. Your follow-up sequence just stops reaching people, and you hear about it when a lead says they never got anything. Even large senders stop halfway: in a 2025 European Commission study, 89% of the high-volume EU email domains measured had a DMARC record, but only 44% had set it to enforce anything.

A sequence adds a second problem, which is time. The email that should go out three days after someone fills in the quiz needs something that holds onto it for three days and still sends it after the server restarts or you ship a change. A web page or a serverless function doesn't wait three days on its own, so you need a scheduled job or a queue, and someone has to notice when it stops. Branch the sequence on whether a lead opened or clicked, and you're building open and click tracking as well.

The Sync to Your CRM That Has to Keep Working

Sending each new lead to HubSpot starts with creating an app or an access token in HubSpot with the right permissions, then mapping your quiz answers onto HubSpot's fields. Connect through HubSpot's standard sign-in and its access tokens currently expire after 30 minutes, so your code has to refresh them on its own, every time.

Then the upkeep starts. Someone changes a field or a rule on the CRM side, and syncs start failing. When the CRM rejects a request or is briefly down, your code has to try again later, or that lead never arrives and nothing tells you. Multiply that by every tool you connect, such as your calendar and your payment provider.

What involve.me Maintains Behind Every Quiz Funnel

When you buy software, someone else is the software company, and plenty of small businesses count on that. According to the same UK government survey, 40% of UK micro businesses say security isn't a major concern when they buy software, because they buy from established companies. That's a sensible instinct. It only holds if someone established is looking after what's behind the screen, and it disappears the day you build the tool yourself.

Every row in that table is something involve.me has already built. We've been building it in Vienna since 2019, and the quiz builder is only the part people see in a demo. The parts we'd least like to rebuild are the ones underneath it, and keeping them working takes a team every week. We're proud of that part, and you'll rarely notice it, which is the point. The short version: trusted by 4,500+ businesses, SOC 2 Type II audited, and GDPR-compliant. The longer version is on our trust center:

• Data hosted on AWS in Frankfurt, with backups held in Germany

• Encryption in transit with TLS 1.2 or higher, and at rest with AES-256

• Penetration testing, and disaster recovery plans that get tested

• A data processing agreement, and a commitment to tell you about a breach within 72 hours

• A public status page with uptime history and incidents

If you'd like the product overview before the plumbing, here's what involve.me is and does in one page.

One Contact Record for the Quiz, the CRM and the Email Sequence

Plenty of small businesses don't have a CRM yet. In Germany, about 1 in 8 firms with 2 to 9 staff used CRM software in 2023, according to Eurostat. If yours is one of the other seven, this part matters most.

In a homemade stack, or a stack of separate tools stitched together with Zapier, the quiz, the CRM and the email platform each keep their own copy of a lead, and a bridge copies data between them. Every bridge is one more thing that breaks quietly.

In involve.me, every submission becomes a contact in involve.me's own CRM, with its answers, score and full timeline on one record. Answers, scores and outcomes map to contact properties. A returning lead updates the existing contact rather than creating a duplicate, and segments update themselves as contacts change. Your automation workflows branch on the same data the quiz collected, and built-in email automation sends from that record, so the email already knows the score without anything copying it across. Wait steps and branches on answers, scores, outcomes, opens and clicks are part of the workflow builder, so there's no scheduler to run and no tracking to build. One subscription replaces a quiz funnel builder, a CRM, and an email automation tool, for less than you'd pay for all three separately.

When you do need another tool, involve.me's native integrations connect to HubSpot or Salesforce by signing in and approving access, suggest the field mapping for you, and show the sync status of every submission, with the full error message when one fails.

The Security and Email Rules You Don't Have to Track

On the email side, every workflow email in involve.me carries an unsubscribe link that can't be removed on any plan, and one opt-out stops every workflow for that contact. Accounts whose bounce or complaint rates climb have their sending limited automatically, which protects deliverability for everyone else on the platform. Paid plans can also send through your own SMTP server.

At the front door, involve.me checks email addresses for a valid format and domain on every plan, and from the Grow plan it can block disposable and personal addresses. The CAPTCHA element (Grow and Scale) supports Google reCAPTCHA and Cloudflare Turnstile, and Scale adds one-time-code verification by email or SMS. Free plan data is deleted after 30 days, paid plans keep it until you delete it, Scale and above can set their own retention period, and contacts can be deleted one at a time or in bulk. That's all work you'd otherwise build and keep up yourself, and a demo shows you none of it.

Should You Vibe Code Your Lead Funnel? A Five-Question Test

Answer these before you build anything that collects leads:

1. Will it store personal data?

2. Will it send email to the people who fill it in?

3. Will it sync with your CRM, a calendar or anything else?

4. Will it stay live for longer than one campaign?

5. Is there someone on your team who can read the code and patch it, and who gets an alert when it breaks?

If you answered no to the first four, vibe coding is a fine choice. If you answered yes to any of them and no to the last one, you'd own infrastructure that no one is looking after.

Vibe coding is the right call for a countdown page for a launch, a quiz for an event that ends on Friday, a prototype you'll show to three customers before deciding anything, or an internal tool that never touches customer data. They're short-lived and store little or nothing, and if one breaks, only your team notices. Build them with whatever's fastest. If that's most of what you need, we're probably not the right tool for you, and that's fine.

Tools That Stay on Your Site for Years Need a Platform

The tools worth the most to a business are the ones that stay put:

• An intake form on your contact page. It asks the qualifying questions, uses lead qualification to score each enquiry, and sends it to the right person with the right follow-up.

• A quote calculator on your pricing page. It helps visitors find the right plan or package, and gives your sales team the numbers before the first call.

• A product configurator. The visitor designs what they want, a kitchen for example, and sees a custom price for it.

Each of these collects personal data, starts an email sequence and feeds a CRM from the first day, and each stays live for years. That's what a sales funnel builder with the CRM and email automation built in is for. If you're weighing up options, we've also put AI funnel generators compared side by side.

The people filling in your form are trusting you with their phone number, and sometimes with their health. Build the quiz with AI, by all means. I'd just want somebody's name next to the database before the first lead arrives.