Enterprise Vibe Coding: What Vibe Coding Is and How Companies Build Secure Applications With It
By Daniel Weise · Founder & Managing Director, MAKONIS
In short
Vibe coding is software development by prompt: an AI writes the code, a person describes in plain language what they want and judges the result instead of the source code. Andrej Karpathy coined the term in February 2025, and Collins named it Word of the Year 2025. It is ideal for prototypes and risky for business software: without guardrails you get vibe coding debt, security holes and applications without a role model. Enterprise vibe coding keeps the AI speed but enforces specification, a DevSecOps pipeline, roles and permissions, a GDPR-compliant architecture and SLA-backed operations – the basis for replacing bloated SaaS tools and modernising legacy systems.
01
What is vibe coding? Definition and origin
Vibe coding is a way of building software in which people describe in natural language what an application should do, and an AI model writes, runs and fixes the code. The person steers with descriptions, screenshots and error messages – and checks whether the application behaves correctly, not what the code looks like.
The term was coined by Andrej Karpathy, founding member of OpenAI and former head of AI at Tesla, in early February 2025 in a post on X. You “fully give in to the vibes” and “forget that the code even exists”. He always hits “Accept All”, no longer reads the diffs and pastes error messages back into the chat without comment. His own verdict: “not too bad for throwaway weekend projects”.
The term spread worldwide within months. Collins Dictionary chose “vibe coding” as its Word of the Year 2025 and defines it as the use of artificial intelligence prompted by natural language to assist with the writing of computer code. Since then the word stands for two things at once: the democratisation of software development – and code that nobody understands any more.
Feb 2025
Andrej Karpathy coins the term on X
2025
Collins Dictionary Word of the Year
45%
of AI code samples with an OWASP Top 10 flaw (Veracode 2025)
02
Vibe coding in 2025/2026: tools and flavours
Three classes of tools shape the trend. Prompt-to-app platforms such as Lovable, Bolt or Replit turn a description into a running web app including database and hosting. AI editors such as Cursor, Windsurf or GitHub Copilot work directly in the code. Terminal coding agents such as Claude Code, OpenAI Codex or OpenCode complete whole tasks on their own, run tests and open pull requests.
In 2026 these agents often work for hours without intervention and build in one pass what used to take a team a sprint. That blurs the line between vibe coding and professional AI-assisted development. Developer Simon Willison draws it clearly: if you review, test and understand every line a language model writes, you are not vibe coding – you are using the AI as a typing assistant.
For companies this distinction matters. The table shows how the three ways of working differ.
| Vibe coding (classic) | AI-assisted development | Enterprise vibe coding | |
|---|---|---|---|
| Who builds | business users, founders, developers in prototype mode | developers | business teams and engineering together |
| Code review | none – “Accept All” | developer reviews every change | automated checks plus a human review gate before every merge |
| Security | chance | depends on the individual | enforced by the pipeline: SAST, SCA, secret scanning, SBOM |
| Roles and permissions | usually none, or only in the UI | depends on team experience | single sign-on, server-side roles, row-level security |
| Operations | unclear | existing IT operations | SLA, monitoring, patches, named owners |
| Goal | prototype, weekend project | more productivity in the team | production business application in days to a few weeks |
03
What is enterprise vibe coding?
Enterprise vibe coding combines the speed of vibe coding with the guardrails business software needs: specification, access control, automated security checks, traceable approvals and managed operations. The AI still writes most of the code – but not in a black box: in your Git repository, in isolated environments and through the same pipeline as any other software.
The difference is not the tool but the system around it. Three questions decide whether an AI-built application may go into production: who may see and change what? What has been checked before it goes live? And who is responsible when the application fails at night?
Why this is urgent now: business teams are already building. Sales, purchasing and HR create their own tools with prompt-to-app platforms – often with customer data and without IT knowing. A ban rarely stops that. A sanctioned path on which business teams build fast and IT stays in control works better.
04
The risks: vibe coding debt, security holes and missing RBAC
AI-generated code often works first time – that does not make it secure. In 2025 Veracode had more than 100 language models solve 80 coding tasks: in 45 percent of cases the models chose the insecure option and introduced a vulnerability from the OWASP Top 10, and for cross-site scripting they failed in 86 percent of relevant cases. Newer and larger models were no more secure.
The typical risks can be named clearly – and controlled with fixed guardrails:
| Risk | How it shows | Guardrail |
|---|---|---|
| Vibe coding debt | duplicated logic, no tests, “organically grown” architecture, nobody understands the code | specification and architecture rules, tests as acceptance criteria, regular refactoring |
| Classic vulnerabilities | SQL injection, cross-site scripting, missing input validation | SAST and DAST in the pipeline, review gate before merge |
| Secrets in code | API keys in the frontend or the repository | secret scanning, Azure Key Vault, short-lived tokens |
| Missing RBAC | every signed-in user sees everything, permissions checked only in the UI | single sign-on via Entra ID, server-side roles, row-level security |
| Open databases | tables without access rules, publicly reachable endpoints | default deny, private endpoints, penetration test before go-live |
| Hallucinated packages | AI suggests libraries that do not exist, attackers register them (slopsquatting) | package allowlist, dependency scanning (SCA), SBOM |
| Agent with production access | agent changes or deletes live data | separate environments, no production credentials in development |
| Compliance gaps | no retention periods, no audit log, data with US providers | GDPR by design, EU hosting, logging |
05
Recognising and paying down vibe coding debt
Vibe coding debt is the technical debt that builds up when AI-generated code is accepted without anyone understanding it. It grows faster than classic technical debt because AI produces more code in minutes than a team can review in weeks. Typical signs: every change breaks something else, the same logic exists in several places, and the agent “fixes” bugs by adding new special cases.
The good news: an AI-built app does not have to be thrown away. A structured vibe coding cleanup turns the prototype into an application you can operate:
- 1Inventory: which AI-built applications exist, who uses them and what data do they process?
- 2Security screening: static code analysis, dependency checks, a secret scan and a look at sign-in and data access.
- 3Capture behaviour: characterisation tests for the most important flows before anyone refactors.
- 4Straighten the architecture: merge duplicated logic, put data access behind a layer, check permissions server-side.
- 5Bring it into the pipeline: Git repository, review gates, automated builds and deployments via infrastructure as code.
- 6Hand over to operations: monitoring, backups, a patch cycle and a named owner with an SLA.
06
Use case 1: SaaS replacement – retiring bloated US SaaS
Many companies pay per seat for US SaaS tools of which they use only a fraction. Every new hire raises the bill, single sign-on and audit logs often come only with the enterprise plan, and the data sits with a provider subject to the US CLOUD Act – even if the data centre is in Europe.
With enterprise vibe coding, a narrowly scoped tool can be replaced in days to a few weeks by your own application running in your Azure environment or a sovereign EU cloud: with sign-in via Entra ID, exactly your processes and no per-user licence. In our worked example of a booking tool like Calendly with 60 hosts and mandatory SSO, building it yourself costs about half over three years. With ten users it does not pay off – which is why the make-or-buy calculation comes first.
| Tool category | Typical problem with US SaaS | Replacement with enterprise vibe coding |
|---|---|---|
| Appointment booking | price per host, SSO only in the most expensive plan | your own booking page with Microsoft Graph, Entra ID and retention periods |
| Forms and surveys | personal data with US providers, price by response volume | form app with data stored in the EU |
| Approval workflows | rigid data model, expensive add-on modules | workflow app built exactly around your process |
| Internal catalogues and portals | per-user price for a few features actually used | lean app with AI search across your own documents |
| Better to stay on SaaS | payroll, GoBD-compliant accounting, CRM suites, collaboration | the vendor maintains regulation and network effects better |
07
Use case 2: legacy reverse engineering – modernising old systems
Many companies run business-critical applications written in Delphi, Visual Basic 6, Access, PL/SQL or early Java and .NET versions. The people who built them are retiring, documentation is missing, and every change is a risk.
Coding agents change the starting point fundamentally: they read hundreds of thousands of lines of legacy code and database schemas, reconstruct business rules, data flows and interfaces and turn them into a readable specification. From that specification a cloud-native application is built module by module – protected by tests that capture the old behaviour.
Where the analysis happens matters: the source code of a legacy application often holds decades of company knowledge. We therefore analyse it in isolated environments with sovereignly operated or local models, not through arbitrary cloud APIs.
- 1Read code and database: make structure, dependencies and dead code paths visible.
- 2Business logic as a specification: rules, calculations and special cases in plain language, confirmed by domain experts.
- 3Characterisation tests: make today’s behaviour automatically verifiable.
- 4Strangler fig replacement: new modules take over step by step while the legacy application shrinks.
- 5Parallel run and shutdown: compare results, then retire the old system in a controlled way.
08
The enterprise vibe coding pipeline: eight guardrails
To keep AI speed from turning into AI risk, every MAKONIS project runs through the same eight guardrails – whether a business team builds a new tool or a legacy system is replaced.
- 1Specification before prompt: business requirements, data model and roles are captured as context for the agents (context engineering).
- 2Isolated workspaces: agents work in sandboxes without access to production data or production credentials.
- 3Approved models: access to language models only through a gateway such as LLMrouter.eu – with EU or local models, budgets and logging.
- 4Git and review gate: every change arrives as a pull request and a person with business responsibility approves it.
- 5Automated security checks: SAST, dependency and licence checks, secret scanning, container scanning and DAST in every pipeline.
- 6Identity and RBAC: sign-in via Microsoft Entra ID or Keycloak with OpenID Connect, roles checked server-side, row-level security in the database.
- 7Traceability: an SBOM per release, audit log, infrastructure as code and documented data flows for GDPR and NIS2.
- 8Operations with an SLA: monitoring, backups, patch management and round-the-clock on-call.
09
GDPR, NIS2, EU AI Act and Cyber Resilience Act
AI-built applications are subject to the same rules as any other software – they just become relevant sooner because they are built faster. The overview shows what the key regulations mean for enterprise vibe coding.
| Regulation | What it means for vibe coding | Guardrail |
|---|---|---|
| GDPR | data protection by design (Art. 25) and security of processing (Art. 32) apply to internal tools too | EU hosting, roles, retention periods, data processing agreements |
| NIS2 | companies in scope must demonstrate security in the development and maintenance of their systems as well | documented pipeline, vulnerability management, SBOM |
| EU AI Act | coding assistants themselves are usually not high-risk; the app you build can be, for example in recruiting. The AI literacy duty applies since 2 February 2025 | classification per use case, training for business teams |
| Cyber Resilience Act | manufacturers of products with digital elements must report actively exploited vulnerabilities since 11 September 2026; the remaining obligations apply from 11 December 2027 | relevant as soon as you sell software: security by design, update process, SBOM |
10
Sources
The figures in this guide are based on the following sources (accessed November 2026):
- Andrej Karpathy: post on “vibe coding” on X (February 2025): x.com/karpathy
- Collins Dictionary: Word of the Year 2025: The Guardian · BBC
- Simon Willison: Not all AI-assisted programming is vibe coding: simonwillison.net
- Veracode: 2025 GenAI Code Security Report: veracode.com
- Replit agent deletes production database (July 2025): Fast Company · The Register
- OWASP Top 10: owasp.org
- European Commission: Cyber Resilience Act reporting obligations: digital-strategy.ec.europa.eu
11
The MAKONIS bridge: AI speed with DevSecOps, GDPR architecture and 24/7 operations
MAKONIS brings both worlds together – we develop with AI coding agents at the speed of vibe coding and deliver software your IT security, data protection and operations teams can sign off.
Build fast: AI software engineering with context engineering, coding agents and review gates – for SaaS replacement and legacy modernisation. With Scalor.io, our self-hosted platform for AI-assisted app development, business teams and developers build apps together inside your own environment, in your Git repositories and through your approval processes.
Ship hardened: DevSecOps pipelines with SAST, dependency scanning, secret scanning, SBOM and infrastructure as code – all the way into your Azure landing zone or a sovereign EU cloud.
GDPR architecture: data stored in the EU, single sign-on via Entra ID, roles and retention periods from day one; language models via LLMrouter.eu or run locally.
Run 24/7: SLA-backed operations from our Managed Service Center in Krakow, a joint venture with aplicado – EU nearshoring instead of offshore, with contacts in Munich. We take over existing AI-built apps as Managed Vibe Coding from €490 per month, starting with an initial screening.
MAKONIS is certified to ISO 27001 and ISO 9001 by TÜV Rheinland and TISAX-assessed. Start with a free 30-minute call with our engineers.
24/7
SLA-backed operations from the Managed Service Center Krakow
from €490
per month: Managed Vibe Coding for existing AI apps
ISO 27001 · TISAX
certified and assessed, plus ISO 9001
Frequently asked questions
Vibe coding means you describe to an AI in everyday language what an application should do, and the AI writes the code. You check the result by trying the application, not by reading the code. You report errors back to the AI until everything works.
Keep reading
Guide · 7 min read
Running LLMs on-premises: a guide for companies
Running an LLM locally means executing a language model on your own or dedicated infrastructure so prompts and data never leave the company. It requires a suitable open-weight model, enough GPU memory, inference software, a connection to company knowledge (RAG) and access control with logging.
Read moreGuide · 7 min read
AI GDPR checklist and the EU AI Act for mid-sized companies
For GDPR-compliant AI, companies need a legal basis, transparency towards data subjects, a data processing agreement with the AI provider, control over data location and retention, and a data protection impact assessment where risk is high. The EU AI Act adds risk-based obligations; for most mid-sized use cases AI literacy of staff and transparency duties matter most. This article is not legal advice.
Read moreGuide · 9 min read
Replacing SaaS with AI: when building your own pays off – the Calendly example
Replacing SaaS with AI means swapping a narrowly scoped SaaS tool for your own application, built with AI coding agents and running in your own cloud environment. It pays off mainly with per-user pricing, a limited feature set and high security requirements. In our example of a booking tool like Calendly with 60 hosts and mandatory SSO, building your own costs about half over three years – with 10 users it does not pay off.
Read moreInnovation from Munich.Active worldwide.
Talk to us about your cloud, DevOps and AI initiatives — no obligation, directly with our engineers.