Vibe Coding in SMEs: a Security Checklist for Non-Developers
You built a small tool with an AI. It reads an Excel list, writes a report and saves you an hour every week. It works. Now you are asking the right question: is it safe enough to use every day?
Anyone working with a coding agent gets a working program in minutes. Whether it is also secure is not visible from the outside. That is true for professionals and beginners alike.
This post explains what "vibe coding" is, where the real risk lies, and twelve questions you can ask without being able to program. The list fits into a coffee break and still catches the typical mistakes.
What vibe coding is and where the problem lies
Vibe coding means describing an app to an AI in plain language and shipping whatever comes out. You do not read the code. You only check whether the result on screen looks right. The tool is usually a coding agent: an AI program that writes, runs and fixes code on its own.
For an internal tool on your own laptop this is often perfectly fine. A script (a small program that does one job automatically) that merges three Excel lists cannot do much harm. At worst a number is wrong, and you notice when you read the result.
It gets risky in two situations. First, when the tool can be reached from the internet. Second, when it handles customer data or other personal data. Then "it works" is no longer enough. It also has to behave correctly when someone does something you did not plan for.
What the recent reports show
In May 2026 Computerwoche reported on research by the security firm Red Access. It found more than 5,000 web apps built with vibe-coding platforms such as Lovable, Replit, Base44 and Netlify that had basic security or login flaws. In many cases, knowing the app's address was enough to see its content. Around 40 percent of these apps contained sensitive information, including medical data, financial documents and customer conversations. Replit and Lovable replied that configuring the app is the user's responsibility.
In June 2026 it-daily published an analysis of 50 public projects built with AI coding tools. It comes from the founder of a security company, so it is not an independent study. The pattern is still instructive: 36 percent of the projects had access keys or passwords directly in the code. The author describes most gaps as errors caused by missing code. Nothing was written wrongly; the protection simply was never requested.
That is the core point. The AI builds what you describe. It reliably adds a login, access restrictions or input checks only when you ask for them.
The checklist: twelve questions before day one
Ask yourself these questions, or ask the coding agent directly. If you cannot answer one, that is already the most important finding.
- Who can reach the tool? Only your computer, the company network, or the public internet? Anything public needs a login.
- Where does the data live? On your laptop, a company server or a cloud service? In which country?
- Does personal data go to a third-party service? Customer data does not belong with a provider you have no data processing contract with.
- What happens with a wrong input? An empty file, a renamed column, a comma instead of a decimal point. The tool should stop and say so instead of silently carrying on.
- Is there a test? A test is a small checking program that runs the tool on sample data and confirms the expected result. Without one, you notice a broken change only when the error reaches a customer.
- Are passwords or keys in the code? Credentials belong in a separate settings file that is never shared, not in the program text.
- What may the tool delete or send? A draft that a person reviews is harmless. An e-mail that goes to customers automatically is not.
- Can it be switched off? Do you know how to stop it if it behaves oddly? Does your deputy?
- Is there a backup? Of the data the tool changes, and of the tool itself.
- Who will understand it in six months? Is there one page of documentation: what it does, where the data comes from, how to start it?
- Who owns it? A name, not "IT". That person decides on changes and notices when something is off.
- What if the AI platform changes or disappears? Does the tool keep running if you cancel the subscription? A finished script usually runs without AI; an app on someone else's platform may not.
How to work through the questions with the coding agent
You do not have to search the code yourself. The coding agent can explain its own code if you ask precisely. A prompt (the instruction you give the AI) could look like this:
Please review this project like a security officer.
Answer in plain sentences for a non-developer:
1. From where can the program be reached?
2. Which data does it read and write, and where to?
3. Are there passwords or keys anywhere in the code?
4. What happens if the input file is empty or a column is missing?
5. Which tests exist, and do they pass?
Do not change anything yet. List the problems first.
You get a list back. Typical items: "the mail service key is on line 12" or "if a column is missing, the program continues with zero". Then ask the agent to fix each item one by one and to write a test that checks exactly that case.
Do not trust the answer blindly. The AI is a very fast, very well-read intern: great for drafts, but check everything that has to be right. Try the case yourself: delete a column in the test file and see whether the tool really stops.
When the checklist is not enough
These twelve questions cover an internal tool used by one person or a small team well. They do not replace a professional review if any of the following applies:
- Several people log in with their own accounts.
- Customers or partners can reach the tool over the internet.
- It processes sensitive data such as health data.
- It is connected to your accounting system, ERP or another core system.
At that point the tool becomes a software project. That is not a failure; it shows the idea works. For the technical side I wrote a detailed checklist for reviewing AI-generated code on the Halvic blog. You can also hand it to a developer who reviews the tool for you.
What you can do today
- List every home-built tool and AI app in the company, with the name of the person responsible.
- Mark every tool that is reachable from the internet or handles personal data. Review those first.
- Give the coding agent the review prompt above and read the answer together with a second person.
- Test one failure case yourself: empty file, wrong column, duplicate entry.
- Write half a page of documentation per tool: purpose, data source, how to start, how to stop.
Frequently asked questions
Is vibe coding simply too risky for SMEs?
No. For internal tools that run on your own computer with your own data, the risk is usually small and the benefit large. It becomes risky when a tool goes online without review or handles customer data. That is what the checklist is for.
Do I need to program to use the checklist?
No. All twelve questions can be answered with common sense or passed on to the coding agent. What matters is that you ask and then try the answers out on a real case.
Is it enough if the AI reviews its own code?
As a first step it is very useful, because the AI finds many typical mistakes. It does not replace your own test with real failure cases. Tools with customer access or many users also need a professional.
What about platforms that publish the app online right away?
There the app is often public with one click. Before publishing, check that only logged-in people can see it and what data it contains. The Red Access findings show that this is exactly how thousands of apps ended up open on the web.
Next step
In the KI-Werkstatt, this checklist is the content of week four, "secure and hand over": every tool gets tests, guardrails (fixed rules about what the tool may and may not do) and a short piece of documentation before it goes into daily use. If you would like to find out first whether it fits your team, book a free first call.
Sources
- Computerwoche: "Sicherheitslücke Vibe-Coding", 11 May 2026, on the Red Access report. https://www.computerwoche.de/article/4169486/sicherheitslucke-vibe-coding.html (accessed 2026-10-01)
- it-daily: "Vibe Coding: Der blinde Fleck der KI", Jonathan Demir, 29 June 2026. https://www.it-daily.net/it-management/ki/vibe-coding-blinde-fleck-der-ki (accessed 2026-10-01)