▶ 7:017 min read

Vibe coding has its blind spots.

Describing an application in natural language and getting a working CRM in a few minutes is impressive. But an application that runs isn't necessarily a secure application, and the gap between the two can be very costly.

Amory Rouault
Amory Rouault · Freelance AI engineer

// in brief

  • AI does what you ask it to. It doesn't flag the risk you didn't think of.
  • The flaws are real: exposed databases, public API keys, APIs without access control.
  • A single question decides: if the tool breaks or leaks, is it serious?
  • Engineers use AI too, then review every generated line.

Vibe coding is impressive. You describe an application in natural language, you press Enter, and in a few minutes you have a working CRM. Tools like Lovable or Bolt have made this accessible to everyone.

The problem: "it runs" doesn't mean "it's secure". And the difference between the two can be costly, very costly.

Thousands of vibe-coded applications have exposed user data directly on the internet. Founders have received a €2,000 AI bill overnight, because their API keys were lying around in a public repository. These aren't theoretical scenarios.

In this article, I go over what I say in the video: why vibe coding is dangerous for critical applications, when it remains acceptable, and what AI won't do for you, whatever the quality of the model.

What AI really brings to development

Let's be precise about what vibe coding has changed. AI has made code generation accessible to non-developers. It's a real, useful and often impressive capability.

What it hasn't brought:

  • Software engineering concepts: patterns, architecture, technical debt.
  • Application security: authentication, authorization, encryption, attack surface.
  • An understanding of infrastructure: networking, service exposure, databases reachable from the internet.
  • The maintainability reflex: what happens in 18 months, when the application has grown?

AI does what it's asked. Ask it to secure personal data, it secures personal data. But if the database that holds it is exposed on the internet without restriction, it won't tell you: you didn't ask it the question.

That's the heart of the problem.

The car analogy

Imagine AI had taught everyone to release a clutch without stalling, turn a steering wheel and change gears. Does that mean everyone can take the highway?

Clearly not. Mastering the basic mechanics of a car doesn't replace years of driving, handling emergency situations, reading traffic.

Vibe coding is exactly that: a barrier to entry that has collapsed, without the fundamental knowledge following.

The real flaws of vibe coding

The risk is concrete. Here's what I actually come across.

Databases exposed on the internet

A typical case: someone asks AI to secure personal data. AI encrypts the sensitive fields. Mission accomplished, on the surface.

Except that the database remains directly reachable from the internet, without authentication. The data is encrypted, but anyone can connect to the database and read everything it contains.

Without a grounding in network security, you don't even know you need to ask AI this question. It's a total blind spot.

API keys in public repositories

Recently, two founders told on Instagram how their Claude API keys had been stolen. Their code was public on GitHub and contained the keys in plain text.

€2,000of API usage in a few hours · keys left in the code

It's a basic mistake: a secret has no business being in the code. It's precisely the kind of mistake an engineering education avoids. And AI won't stop you if you didn't think to ask it the question.

Unprotected APIs

An API generated by vibe coding exposes data without properly checking permissions. Anyone with the right URL can query other users' data.

This flaw has a name, Broken Object Level Authorization (BOLA), and has topped the OWASP top 10 for years. AI doesn't anticipate it on its own if you don't know the risk exists.

The false sense of control

The subtlest danger of vibe coding: the result is impressive, so you think you are in control.

You launch Lovable, you describe your application, you get a clean interface, data that gets saved, user accounts that work. The interface looks like the real thing. The feeling of competence is real.

But the apparent quality of the interface says nothing about what's happening underneath.

  • Are sessions handled correctly?
  • Do requests go through server-side validation?
  • Are uploaded files isolated per user?

These questions have no visible answer from the interface. Only someone who has the concepts in mind thinks to ask them.

The right question: if it breaks, is it serious?

The question to ask is less about vibe coding itself than about the tool: if it breaks or if it leaks, is that a problem?

The cases where it works

  • An internal dashboard that pulls in Google Analytics KPIs.
  • A tool that automates a repetitive task, with no sensitive data.
  • A prototype to test a concept before really investing.
  • A throwaway script, for one-off use by a single person.

In these cases, the consequences of a bug or a flaw remain limited. Vibe coding is a legitimate accelerator.

The cases where it breaks down

  • A CRM with client data: contact details, history, contracts.
  • A SaaS application that collects user data.
  • A tool at the heart of a company's sales activity.
  • Any application that touches payments, health or personal data.

Here, an incident can have serious consequences: GDPR, loss of client trust, business interruption, legal fees.

Every system eventually fails

Security is one risk among others. There is also what happens when the application grows.

Everything fails, all the time. Werner Vogels, CTO of Amazon for more than 20 years

The principle: any system that can fail will eventually fail. It's a law, not a possibility. A web application, a CRM, a SaaS: anything can crash. Maybe in 6 months, maybe in 2 years. But it will happen.

The concrete scenario

A web agency owner develops his CRM in-house with vibe coding. A good idea on the surface: a custom tool, adapted to his processes. But:

  • No dedicated developer maintains the tool.
  • Over 12 months, the application has grown as features were added.
  • One day, it crashes. Or worse, it behaves strangely with the data.

Two scenarios. With someone competent, in-house or external, who steps in quickly: it's an interruption, not a crisis. Without: you stop everything to debug a vibe-coded tool you don't really master, while your business runs at a crawl.

For an agency that makes several tens or hundreds of thousands of euros per month, spending 3 days debugging a CRM is a crisis.

Engineers use AI, then review

A point to put things back in perspective: engineers at Stripe and other solid tech companies also generate code with AI. Vibe coding exists in these teams.

Only, you don't see it. And that's precisely the sign that it works.

An experienced engineer reviews the generated code with their own knowledge. They identify the flaws, fix the bad patterns, adapt the architecture. AI produces the code, the human validates and fixes.

The final result erases AI: it did take part, but the human review is solid enough for the code to be reliable.

That's the real use of AI in development. Not "AI coded it, we deploy".

What I check on a vibe-coded tool

If you already have a vibe-coded application running, there's no need to throw everything away. What matters is knowing where you really stand. The points to check:

  • Database: directly reachable from the internet, or behind a secure access layer?
  • API: does each endpoint check that the logged-in user has the right to access the requested data?
  • Secrets: are API keys and database credentials in the code, or in unversioned environment variables?
  • Authentication: do sessions rely on proven standards, or on something AI made up?
  • Dependencies: are the libraries used up to date and free of known vulnerabilities?
on audit.application
01 database.exposure → behind an access layer?
02 api.endpoints → permissions checked on every request?
03 secrets → out of the code, unversioned?
04 auth.sessions → proven standards?
05 dependencies → up to date, no known flaw?
✓ flaws mapped · fixes prioritized

If you can't answer these questions for your application, that's the first sign that an outside technical review would be useful.

The laws of engineering stay the same

Vibe coding has changed something real in access to development. For some uses, it's a good thing.

But an exposed database remains an exposed database, whether the code was written by a human or by an AI. An API key in a public repository remains an API key in a public repository.

The rule is simple. Vibe coding throwaway tools and light automations: yes. Vibe coding, without technical review, tools that touch client data, payments or the core of your business: no.

Frequently asked questions

No. For internal tools without sensitive data, prototypes or light automations, it's a perfectly valid approach. The risk begins when you apply it to critical tools (client data, core sales activity, personal data) without technical review.

Yes, with a complete technical review. The audit maps the existing flaws, then you fix them one by one. It takes longer than starting from a clean architecture, but it's doable for an application that has business value.

These tools are improving, and some include basic safeguards. But they can't anticipate the context of your application: permissions, network exposure, the data you handle. Secure generic code doesn't replace an architecture designed for your use case.

Probably, the models are progressing fast. But in 2026, we're not there yet: the results are visually impressive, the security and architecture layer remains insufficient for a critical application without competent human review.

The only reliable way is a technical audit by someone who knows the common flaws of web applications: network exposure, permission management, secrets, vulnerable dependencies. An outside review with the right tools concretely identifies what is causing problems.

Amory Rouault
// written by

Amory Rouault · Freelance AI engineer

I design and code AI agents, automations and custom applications. On YouTube, I share what I learn in the field.

$ amory audit app

Is your vibe-coded application really secure?

Let's talk about your project

// youtube

The next video, straight in your feed.

// contact

Do you want a technical review of your application? Let's talk about it.