Vibe Coding: What Non-Developers Can Safely Build and When Your App Needs Experts
Vibe coding means you let an AI build software without reviewing the code. For prototypes and small tools without real data, that is perfectly fine. Once customer data, payments, or other people come into play, the app needs an expert review before launch.
You can build an app in an afternoon without understanding a single line of code. That part really works now. The trouble rarely starts with the build. It starts on the day real data goes in. The app looks just as finished as before, even if any stranger can reach the database. This guide sorts apps by risk and gives you four tests you can run yourself, no coding skills required.
What exactly is vibe coding?
Vibe coding is building software with a language model without reviewing the code it generates. Programmer Andrej Karpathy coined the term in February 2025. The sharpest definition comes from developer Simon Willison: if you review the code and can explain it, you are doing normal software development with AI help. So the label does not depend on the tool. The same app in Lovable or Claude Code is vibe coding or it isn't, depending on whether someone understands the result and takes responsibility for it.
Willison sees a real gain here, because people without programming experience can build their own tools. His condition: bugs and security holes must not be able to cause serious harm.
What can you safely build without coding skills?
Anything that works without real data or only affects you. What decides it is not the size of the app but the damage if it sits open on the internet. Our take, in three levels:
| Level | Examples | What to do |
|---|---|---|
| Green: no harm possible | Click-through prototype with sample data, pitch demo, calculator using public numbers | Go ahead. Enter no real data, not even "just for a quick test" |
| Yellow: harm stays in-house | Internal team tool with company data but no sensitive personal data | Login required, run the four tests below, have someone experienced look at it before rollout |
| Red: harm hits other people | Customer data, HR data, health data, payments, write access to your CRM | Do not go live without an expert review. Vibe coding works for the prototype here, not for production |
Watch for the moment a yellow tool turns red: as soon as someone imports a customer list, the new level applies, even if nothing in the code has changed. For how to turn a prototype into a solid product, see our article on MVP development and its limits.
How insecure are these apps really?
Often far less secure than they look, but the numbers going around measure very different things. Many articles list them side by side as if they were about the same thing:
| Number | What it actually measures | Limitation |
|---|---|---|
| 91 percent | 182 of 200 randomly sampled, publicly running apps have at least one vulnerability of any severity | Only open source apps from GitHub, built with Claude Code or Lovable. No control group of hand-built apps. Preprint with no named peer review venue |
| 1 in 5 | Share of organizations on such platforms that are affected, according to Wiz (September 2025) | Refers to organizations, not apps. Wiz does not give a sample |
| Around 40 percent | Share with private data among the more than 5,000 apps without basic protection that RedAccess found (summary of a Wired report) | The base is only apps that were already unprotected, not all apps. The vendors dispute the timeline and individual findings, but not that the apps were open |
The first study, a preprint by Deng, Fan, and Meng (version of September 14, 2026), holds something more useful for you than the 91 percent: of 1,186 vulnerabilities, just under 66 percent were rated critical or high. The largest group, with 339 cases, is broken access control, meaning users can reach data that isn't theirs. Next come injection (commands smuggled in through input fields) and login flaws.
Moltbook, a platform for AI agents, shows what this looks like in a real incident. Its founder said he had not written a line of code himself. In January 2026, Wiz reported a database anyone could read and change: 1.5 million API tokens (access keys) belonging to agents, 35,000 email addresses, and 4,060 private conversations between agents. The cause was a missing setting, Row Level Security (RLS). In the Supabase database, it controls who may see which row.
Why does the AI build in these holes?
Our take: because it optimizes for what you see and ask for. The preprint describes eight recurring failure patterns. Five of them are worth knowing as a non-developer:
- The fix weakens the protection: You report a bug, the agent loosens a security rule, and the bug is gone. To you, that looks like a success.
- Demo before security: The agent builds an insecure shortcut for the quick demo.
- Only half changed: A security rule applies in some places but not in others.
- Not requested, not built: Protection you didn't explicitly name is missing.
- Your step, and you don't know it: The agent leaves a step for secure operation to you, and you miss it.
Better prompts help, but they aren't enough: according to the authors, frequency drops while the underlying risk remains. In practice, a security instruction in your prompt does not replace checking the result.
Which four tests can you run without coding skills?
These four tests target the largest vulnerability group and typical Wiz findings. They don't prove security, they only find big holes. If you find one, the app isn't ready to launch. Run them only on your own app and with test accounts.
- Incognito test. Open a private browser window, stay logged out, and go straight to the addresses of your internal pages. If you see content, the protection is missing. Internal apps without a login are among the four risks Wiz describes.
- Two-account test. Create test accounts A and B and use A to create a record. Copy its address, log in as B, and open it. If you see A's data, you have the problem of the largest vulnerability group. As B, also try changing the record, and test the admin pages with a normal account.
- Key test. Right-click, view the page source, then search for the names of your services, such as OpenAI or Stripe, and for words like "secret" or "password". Secret keys have no place in the browser. A Supabase publishable key may appear there, because the protection then sits in the database rules. Many scripts live in separate files, though, so this test only finds the obvious.
- Table test. In Supabase, check whether RLS is active on every table and what the rules allow. Supabase calls for RLS on every table in the schema exposed through the API. A rule that lets everyone do everything is as bad as no rule.
The most important habit: repeat the tests after every fix the agent builds, even if the fix had nothing to do with security. That follows directly from the first failure pattern above.
Are Lovable's security features enough?
No. They are a good first hurdle, but not a green light. Lovable scans automatically at publish, for example for tables without access control and rules that let everyone through. On top of that, there is a deeper scan you have to start yourself, and its findings aren't fixed automatically. Lovable itself says this doesn't replace a thorough security review, that the responsibility is yours, and that apps with sensitive data should also get a professional review.
The platform itself can have flaws too. Lovable admitted in April 2026 that from February 3 to April 20, chat histories and source code of public projects were accessible to other users. According to Lovable, private projects and Lovable Cloud were not affected, and all projects public at the time, except official templates, were then set to private. Our advice from this: never put customer data or keys in the chat, because the chat history is a data store too. For how to protect this technically in Lovable, see our article on Lovable webhooks and Events & Hooks.
When does your app need experts?
Our rule of thumb: as soon as you answer yes to one of these questions, plan an expert review before launch.
- Does it hold data from people who entrusted you with that data?
- Does money flow, or does the app write to systems like your CRM or accounting software?
- Would your business stall if the app went down for a day?
- Do people outside your team use it?
- Is there nobody on your side who can explain who may see which data and why?
Operations often get forgotten here. Who takes care of it when the app is down? Who updates it, and who gets the error message? With Lovable, even an empty credit balance can stop your live app. Settle this before other people rely on the app.
That doesn't mean you have to throw the prototype away. Our take: a vibe-coded prototype often shows better than any requirements document what the business team really wants. If you'd like a second opinion on that, talk to us.
What does this mean for your company?
Our take: banning it achieves little, because the apps get built anyway, just without your knowledge. Three steps for this week make more sense:
- Take inventory. Ask your team who has built what with Lovable, Replit, Claude Code, and similar tools, and where it runs. Wiz recommends this too. According to the summary cited above, RedAccess also found strategy papers and chatbot logs among the open apps.
- Classify. Every app gets green, yellow, or red from the table above, plus one responsible person. Red without a review comes offline until the review is done. A login in front of the interface isn't enough if the database itself is open.
- Set a one-sentence rule. For example: no real personal data in apps without reviewed access rules. Everyone understands that, even without training.
A real incident also comes with a deadline: if personal data was openly accessible, the company must notify the supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware of it under Art. 33 GDPR, unless a risk to the people affected is unlikely. Whether that applies is best settled right away with your data protection advisor. "We only built it quickly" changes nothing about this duty.
Does your app handle real data? Write to us, and we'll take a look at it before launch.
Frequently asked questions
Is my app secure once it has a login page?
Not if the check only runs in the browser. Wiz describes apps where the password could be found in the JavaScript and a value set by hand in browser storage bypassed the login. The check belongs on the server.
Which tools are used for vibe coding?
The studies and reports cited here mainly mention Lovable, Replit, Base44, and Claude Code. For the risk, the tool matters less than whether someone reviews the result.
Does it help to ask the AI to make the app secure?
It helps somewhat, but it should never be the only measure. In the experiment in the preprint by Deng, Fan, and Meng, adding "ready for production" to the prompt cut the share of runs in which a known vulnerability came back from 25.7 to 11.0 percent. No tested variant removed all vulnerabilities.
Can I use real customer data for testing?
You shouldn't. Sample data is enough for almost any test, and real data pushes your app straight to the highest risk level. That also applies to the chat with the AI tool.
What should I do if I find a hole in a live app?
Take the app offline, lock database access, replace any affected keys, and check with your data protection advisor whether you have to report it.
Sources
- Deng, Fan, Meng: Understanding the (In)Security of Vibe-Coded Applications (arXiv preprint)
- Wiz Research: Common Security Risks in Vibe-Coded Apps
- Lovable: Security documentation
- Lovable: Our response to the April 2026 incident
- Supabase: Row Level Security
- Simon Willison: Not all AI-assisted programming is vibe coding
Tell us what you're planning. You'll get an honest assessment within 24 hours.
More articles