Originally published on shipfastdev.com.
Vibe coding, building an app by describing it to an AI tool instead of writing the code, is a fast and cheap way to get a working prototype in front of users. It breaks down in three places: security (especially who can read and write which data), maintainability once the code grows, and running a production system with real users. The right moment to bring in a developer is before you store real personal data, take payments, or find that every new prompt breaks something that used to work.
What vibe coding is
Collins Dictionary made "vibe coding" its Word of the Year for 2025 and defines it as "the use of artificial intelligence prompted by natural language to write computer code." The term was coined by AI researcher Andrej Karpathy. In practice it covers AI app builders that generate a full application from a chat, and coding agents that edit a codebase on request while the person mostly checks the result in the browser.
Professional developers use AI heavily, but mostly not this way. In the 2025 Stack Overflow Developer Survey, 72.2% of respondents said vibe coding was not part of their professional work.
What vibe coding does well
- Testing an idea with real people. A clickable, working prototype gets better feedback in user interviews than a slide deck.
- Demos and landing pages. Showing investors or early customers something that runs.
- Internal tools used by a handful of trusted people, with no sensitive data.
- Clarifying requirements. A prototype is an excellent specification. A developer can scope the real build from it much faster than from a written brief. Our guide on how to scope an MVP explains what to capture.
The common thread: the cost of throwing it away is low.
Where it breaks
1. Security and data access
The best-documented failure is missing data-access rules. Some AI app builders generate apps on a hosted Postgres database such as Supabase, where the browser can query the database directly and Row Level Security (RLS) policies decide who can see what. Supabase's documentation is blunt: "A table in an exposed schema without RLS is readable and writable by any role with a grant on it."
In 2025, a security researcher reported CVE-2025-48757: insufficient RLS policies in projects generated by Lovable allowed unauthenticated attackers to read or write database tables of generated sites. The advisory lists exposed data including names and email addresses from user tables, API keys for third-party services, and payment and subscription details. Lovable disputed the CVE, arguing that each customer is responsible for protecting their application's data. The researcher's follow-up statement notes that the security scanner Lovable later added only checked whether an RLS policy existed, not whether it was correct.
That dispute is the point: whatever the tool's terms say, if your app leaks customer data, it is your problem. And the code itself carries the usual AI risks: in Veracode's 2026 tests, AI-generated code passed security checks only 56% of the time. Our AI-generated code security checklist covers what to verify.
2. Production data and autonomous agents
In July 2025, SaaStr founder Jason Lemkin was testing Replit's AI agent when it deleted a production database holding data on more than 1,200 executives and over 1,190 companies, even though the project was in a "code and action freeze." Replit's CEO called it "unacceptable and should never be possible" and announced automatic separation of development and production databases, better rollback, and a planning-only mode. The lesson for any project: an AI agent should never have direct write access to production data, and backups must be tested.
3. Maintainability
66% of developers in the Stack Overflow survey named "AI solutions that are almost right, but not quite" as their top frustration with AI tools. In a vibe-coded app nobody has read, "almost right" accumulates: duplicated logic, inconsistent data models, no tests. Each new prompt has to work around what is already there, and changes start breaking unrelated features. That is technical debt, built up faster than usual.
4. Running a real product
A demo needs to work once. A product needs backups, monitoring, database migrations that don't lose data, email that reaches inboxes, payment webhooks that are processed exactly once, and releases that don't take the site down. These are operational decisions, and an AI builder will not make them for you unless you know to ask.
When to bring in a developer
Bring someone in, even for a short review, when any of these become true:
- You are about to store real users' personal data or take payments.
- People pay you and expect the product to be up.
- Each change breaks something else, or you are afraid to touch parts of the app.
- You need integrations that must be reliable: payments, an ERP, a mobile app in the stores.
- A customer, investor or partner asks about security or data protection.
How to hand over a vibe-coded MVP
- Get the code into a Git repository you own. Make sure you, not the tool or a freelancer, control it. See code ownership and IP when outsourcing.
- Inventory and rotate secrets. Any API key that appeared in a prompt, in front-end code or in a shared screenshot should be replaced. Secret keys never belong in the browser.
- Audit data access. Check every table and endpoint: is access restricted, and do the rules match who should see what?
- Add tests on the flows that matter: sign-up, login, payment and anything involving permissions.
- Decide: keep, refactor or rebuild. If the data model is sound, refactoring is usually cheaper. If it isn't, rebuilding the core while reusing the prototype as a specification can be faster. Our article on no-code vs custom development and our MVP cost breakdown help with that decision.
Frequently asked questions
Can you build an MVP with vibe coding?
You can build a prototype that works well enough to test demand and gather feedback. Before it handles real personal data or payments, it needs a security and data-access review by someone who can read the code.
Is vibe coding safe for production apps?
Not without review. Documented incidents include data exposed through missing database access rules (CVE-2025-48757) and an AI agent deleting a production database. AI-generated code also fails a large share of security tests.
When should I hire a developer after vibe coding?
Before storing real user data or taking payments, and as soon as changes start breaking unrelated features. A short review early costs far less than fixing a data leak later.
Should I rebuild my vibe-coded app from scratch?
Not automatically. If the data model and access rules are sound, refactor and add tests. If they aren't, rebuild the core and keep the prototype as a living specification.
Sources
- Collins' Word of the Year 2025: AI meets authenticity as society shifts, Collins Dictionary, 2025-11
- 2025 Developer Survey: AI, Stack Overflow, 2025
- CVE-2025-48757: Lovable insufficient Row Level Security, Matt Palmer, 2025-05
- Statement on CVE-2025-48757, Matt Palmer, 2025-05
- Row Level Security, Supabase Docs, 2026
- AI-powered coding tool wiped out a software company's database in 'catastrophic failure', Fortune, 2025-07
- LLMs Are Getting Smarter but Not Safer: Veracode 2026 GenAI Code Security Report, Veracode, 2026-07
Top comments (0)