A current interview with Ben Sufiani gets to the heart of a topic that is currently occupying many companies: by his own account, Sufiani cannot program and yet builds apps, websites and his own AI agent. He describes how he deals with code that he himself cannot check – and draws a clear conclusion from it: vibe coders are an easy target for attackers.

This assessment is less a warning about AI tools than about a gap in the approach. The path from idea to running application has become short. The path to a robust review, to safeguarding and to regulated operation has remained just as long.

What is meant by vibe coding

Vibe coding refers to a way of working in which an application is created predominantly through descriptions in natural language: you formulate what the program should do, an AI model generates the code, and what is assessed above all is the visible result – does the interface work, is the flow right, does it look good.

That is precisely where the shift lies. What is assessed is the behaviour of the application in the normal case. What is not assessed is what happens in the event of an error, how data is stored, which interfaces are open and which third-party libraries have been included. These questions determine security – and they are not visible on the surface.

Why a lack of review is an operational risk

As long as an AI-generated prototype runs locally and is fed with made-up data, the damage is limited. It becomes problematic from the moment the application is publicly accessible, processes real customer data or is connected to internal systems.

Then the same requirements apply as for any other software in the company: access protection, logging, data protection, updates. An application whose structure nobody in-house can explain is difficult to safeguard and even more difficult to maintain. If something fails or a vulnerability appears in a component that is used, there is no starting point for a response.

There is also an organisational effect: solutions that come about bypassing IT do not appear in any inventory list. Nobody knows which systems exist, where they are hosted and who looks after them when the person who built them leaves the company or changes department.

The wrong reaction would be a ban

Business departments turn to AI tools because they want to solve a concrete problem quickly – an internal tool, a form, a small analysis. This impulse is productive. Anyone who bans it across the board does not get less shadow IT, only less well-documented shadow IT.

A clear separation between experiment and production makes more sense. Prototypes may be fast, messy and AI-generated. Anything that goes into regular operation goes through the same process as any other application.

A workable model for AI prototypes

  • Protected sandbox: a segregated environment for experiments, without access to production data and without public accessibility.
  • Data rule in advance: clearly define which data may enter an AI tool or a prototype at all – especially in the case of personal data.
  • Threshold to production: a defined point from which a technical review is mandatory, for example in the case of external access, real data or connection to existing systems.
  • Review by people with expertise: examination of authentication, permission assignment, input validation, handling of credentials and integrated dependencies.
  • Automated checks: dependency and security scans in the delivery chain so that known vulnerabilities in third-party components are noticed before they go live.
  • Inventory and responsibility: every production application with a named responsible person, hosting location and update path.

None of these points is new. What is new is only that they must now also apply to applications that do not originate from the IT department.

What agencies and IT service providers contribute to this

The role is shifting. In the past the brief was "build us this tool". Today what more often arrives is a finished prototype with the question of whether it can be operated as it is. This is not a step backwards but a usable starting point: the business need has already been clarified, the processes have been worked through, misunderstandings about requirements fall away.

The work therefore shifts to review, hardening and transfer into a maintainable structure. This often includes the honest answer that rebuilding on a solid basis is cheaper than refurbishing a generated draft – especially when the application is to run and be further developed over years.

Maintainability is part of security

Security vulnerabilities rarely arise only at the first deployment. They arise over time because libraries become outdated, dependencies are no longer maintained and nobody installs the updates. An application without a clear structure, without tests and without documented decisions is difficult to update – and therefore remains unupdated.

Anyone who uses AI in the development process therefore needs not only a review at the start but also responsibility for ongoing operation. This applies to generated code just as much as to hand-written code.

For web and software projects this means: AI prototyping belongs in the toolbox, because it makes ideas tangible quickly and shortens coordination.What matters is the line between experiment and production system – anything that processes real data or is publicly accessible needs review, hardening and a named owner.Companies that define this line now gain the speed advantage without quietly shifting the risk into operations.

Sources