Skip to content

An experiment becomes a production tool only after approval.

A review path using the example of a sales director who uses vibe coding to build a tool in one afternoon that sorts open quotes from a CRM export, and tries it out with anonymized data in a separate environment. Before the team keeps using it, it passes five checks: a developer reads the code, replaces a hard-coded API key with an entry in the secrets vault, checks the packages it uses and adds tests, and it is settled who maintains the tool. Only then does it work with real data, and IT takes over maintenance.

In vibe coding, a person describes what a program should do and a language model writes the code. The person runs the result and feeds error messages back to the AI. Nobody reads the code itself.

AI researcher Andrej Karpathy coined the term in a post on X on February 2, 2025. He wrote that he hits “Accept All,” no longer reads the diffs, and pastes error messages back in; for throwaway weekend projects, he said, it's not too bad. Collins Dictionary named “vibe coding” its Word of the Year for 2025.

For executives, the question is where a prototype ends. Once vibe-coded software touches customer data, credentials for company systems, or many users, it has to meet the same bar as any other software: security, data protection, maintenance, and liability.

Chapter 1 of 7

How vibe coding works.

The process is a loop. In a chat or AI coding tool, the person describes what they need, say “a page that reads an Excel file and shows revenue per customer as a table.” The model writes the code, the person runs the program and judges only what they can see. When something breaks, they paste the error back, and the model edits the code.

Tools that create files, run commands, and install packages on their own shorten the loop further, acting like an AI agent inside the project folder. That saves time and widens what happens without review: a terminal command takes effect immediately.

Karpathy named the downside too: the code grows beyond his usual comprehension, and some bugs he works around instead of fixing.

Chapter 2 of 7

Review separates it from AI-assisted development.

Writing code with help from an AI assistant isn't automatically vibe coding. The test is whether someone who can judge code reads, understands, and owns the result. In AI-assisted development, the model suggests code; a developer reviews every change, adds tests, and sends the code through the same checks as code written by hand.

Germany's Federal Office for Information Security (BSI) and France's cybersecurity agency ANSSI recommend exactly that in their joint 2024 paper on AI coding assistants. Generated code should generally be checked and reproduced by the developers. The assistants are no substitute for experienced developers, they write, and unrestrained use can have severe security implications. Teams that code faster with AI also need more quality assurance, the paper says.

Chapter 3 of 7

What unreviewed code can do in production.

The biggest risk is security flaws nobody notices because nobody reads the code. In a user study published at the ACM CCS security conference in 2023, participants with an AI assistant wrote significantly less secure code than those without one. They were also more likely to believe their code was secure. The BSI calls this automation bias: convincingly worded output gets accepted unchecked.

A second risk is made-up packages (software libraries). A study presented at the 2025 USENIX Security Symposium had 16 models generate 576,000 code samples. On average, at least 5.2 percent of the packages named by commercial models didn't exist, and 21.7 percent for open-source models. Attackers can register those names and publish malicious code under them.

Then there are hard-coded API keys, and confidential data that, depending on the contract terms, ends up with the AI provider. The OWASP project lists model output executed as a command without validation as a typical weakness.

Chapter 4 of 7

Liability and data protection don't depend on who wrote the code.

If a tool processes personal data, Article 32 of the EU General Data Protection Regulation requires appropriate technical and organizational measures to secure that processing. If nobody knows what the code does, it's hard to show those measures are in place.

Companies shipping software to EU customers should know the new Product Liability Directive (EU) 2024/2853. It explicitly counts software as a product (Article 4(1)) and applies to products placed on the market or put into service after December 9, 2026 (Article 2). Under it, manufacturers are liable for defective products regardless of fault. Specific cases are for legal counsel to assess.

Internally, someone has to own the tool, even one built in an afternoon, and update it when an interface changes or a library it uses discloses a vulnerability.

Chapter 5 of 7

Approval rules separate experiments from production.

A ban does little good, because the tools are freely available; without an approved option, shadow AI follows. The BSI and ANSSI advise providing approved tools with clear contract terms. A clear line between experiments and production works better.

Experiments run in a separate environment without access to production systems, with test or anonymized data and approved AI tools only. Inside that boundary, people can build fast and without review.

An experiment becomes a production tool only after approval. Someone who can judge code reads it and owns it. There are tests, version control records every change, API keys sit in a secrets vault instead of the code, and every package has been checked. And it's settled who maintains it. If any of this is missing, it stays a prototype. Review depth follows the possible damage: a script that reformats a spreadsheet needs less than an app that places orders.

Chapter 6 of 7

The prototype is cheap; running it is not.

Vibe coding makes a first working draft cheap. The later costs don’t go away. Review, testing, security hardening, and maintenance run for as long as the tool is in use, and anyone taking over the code has to understand it first. The trade-off resembles build versus buy: a tool the company builds belongs to the company, with every obligation attached.

Good use shows when the prototype answers a clear question, such as whether a new report helps the team in daily work. Someone then decides on purpose whether to discard, rebuild, or adopt it after review, through the same approval as any other software. Bad use shows in tools that drift into daily operations unnoticed, work with real data, and have no owner.

Chapter 7 of 7

Where the term stops working

Vibe coding describes a way of working, not a product; the same tools work with careful review. Once a person reads, tests, and takes responsibility for the code, it is AI-assisted development. Vibe coding also differs from low-code and no-code platforms, which assemble applications from prebuilt components, and from AI agents, which act in other systems on their own.

Frequently asked questions

What is vibe coding?

Vibe coding is a way of working in which an AI writes code from plain-language instructions and the person accepts it without review. They only check whether the program runs and feed errors back to the AI. Nobody reads the code itself.

Who coined the term vibe coding?

AI researcher Andrej Karpathy coined it in a post on X on February 2, 2025. He described accepting the AI's changes unread and called it not too bad for throwaway weekend projects.

What are the risks of vibe coding?

The main risks are undetected security flaws, made-up packages that attackers can exploit, hard-coded API keys, and confidential data in prompts. Missing ownership adds risk when a prototype stays in use unnoticed.

Will vibe coding replace developers?

Not for production software. The BSI and ANSSI write that AI coding assistants are no substitute for experienced developers. Architecture, security, testing, and maintenance still need people who can judge code.

Can you build professional software with vibe coding?

Not without review. Professional software needs code review, tests, version control, and settled maintenance; once those steps are added, it is AI-assisted development. A vibe-coded prototype can still be a template a team reviews or rebuilds.

What is the difference between vibe coding and traditional software development?

The difference is whether anyone reads the code and takes responsibility for it. Traditional development treats requirements, review, testing, and documentation as fixed steps; vibe coding skips them. The effort shifts to the point where a prototype should go into production.

    ↑ ↓ select · Enter open · Esc close