- Vibe coding means building software by prompting an AI and accepting the output without reading the code. Andrej Karpathy coined the term in February 2025.
- Martin Fowler calls it best suited to disposable software for a small audience that accepts the risks, not widely used products.
- Veracode’s July 2025 research tested more than 100 language models on 80 coding tasks; in 45 percent of test cases the generated code introduced vulnerabilities classified within the OWASP Top 10.
- The most common production gaps are access control, exposed secrets, missing rate limits and no tests or monitoring, not exotic attacks.
- Most AI-built apps can be hardened rather than rebuilt, if the data model and the split between client and server are sound.
What is vibe coding, and when is it not enough?
Vibe coding is building software by prompting an AI model and running what it produces without reviewing the code. It suits throwaway tools for yourself; it is not enough for software other people trust with their data.
Martin Fowler’s definition of vibe coding is building software by prompting a large language model and trying out the results without reviewing the generated code. He credits Andrej Karpathy with coining the term in a post in February 2025. Fowler also draws the line this guide is about. Vibe-coded software often has problems with maintainability, correctness and security, so he sees it as best for disposable software used by its author or a small group who accept the risks. Software that strangers rely on, with their data, is a different category.
That is the moment most founders reach. The prototype proved the idea, a few customers want it, and the same app is about to hold real accounts, payments or personal data. Nothing about the code changed; the stakes did.
Why do vibe-coded apps run into trouble in production?
Because nobody reviews the generated code. AI models often produce insecure code, and without review those flaws reach production along with the features.
Veracode’s GenAI code security research, published on 30 July 2025, ran 80 curated coding tasks across more than 100 large language models. In 45 percent of all test cases the models introduced vulnerabilities classified within the OWASP Top 10, and when a task allowed a secure or an insecure approach, they chose the insecure one 45 percent of the time. The tasks were individual pieces of code, not whole applications, but the direction is clear: AI-written code needs the same review as human-written code, and vibe coding removes that review.
In our experience the gaps are rarely clever. They are the parts a demo never exercises:
- Database tables any signed-in user, or anyone with the public key, can read or change.
- Secret API keys shipped to the browser or committed to the repository.
- Permission checks done in the user interface but not on the server.
- No limits on sign-ups, logins or calls to third-party AI and email services, so a script can hammer them.
- No tests, so every new prompt can quietly break something that worked yesterday.
- No error tracking or backups, so the first real incident is discovered by a customer.
Risky
Demo-ready
Works for the person who built it, on the happy path, with test data. Security lives in the interface: buttons are hidden, but the API behind them is open.
Safer
Production-ready
Every request is authorized on the server, secrets stay server-side, abuse is rate-limited, data is backed up, failures raise alerts, and tests protect the paths customers depend on.
The vibe coding security checklist
Work through it in order; the first two groups cover the issues that most often expose real customer data. Ticks are saved in your browser only.
0 of 15 done
Ticking a box records your own assessment. It is not proof that a control works; only testing shows that.
Where does each checklist item come from?
Each item comes from a published security standard or official platform documentation, linked below so you can check it yourself.
- Database access: Supabase’s Row Level Security guide says a table in an exposed schema without RLS is readable and writable by any role with a grant on it, and tells you to enable RLS on every table in an exposed schema. If your app uses Supabase, check this first.
- Record-level authorization: Broken Object Level Authorization is the first entry (API1:2023) in the OWASP API Security Top 10, and Broken Function Level Authorization (API5:2023) covers unprotected admin functions.
- Rate limits: Unrestricted Resource Consumption is API4:2023 in the same list.
- Secrets: OWASP’s Secrets Management Cheat Sheet covers storage and rotation, and GitHub documents secret scanning for repositories.
- Logging: OWASP’s Logging Cheat Sheet lists data that should not be written to logs, such as passwords and session tokens.
- Going further: the OWASP Application Security Verification Standard is a detailed requirement list if you need a formal security baseline.
Should you fix your vibe-coded app or rebuild it?
Usually fix it. A rebuild freezes the product while competitors keep shipping, and it throws away what the prototype already proved. This is our engineering judgement, not a rule, so look at three things first:
- Harden it if: the data model fits the business, the server and the browser are clearly separated, and tests can be added around the critical paths.
- Rebuild parts if: one area, often payments, permissions or a tangled back end, cannot be made safe without rewriting it. Replace that area and keep the rest.
- Rebuild if: business rules live in the browser, the data model does not match how the business works, or nobody can build and deploy the code outside the AI tool.
Related reading
If you are still shaping the product, start with the product discovery process. For a SaaS that will serve many customers, see multi-tenant SaaS architecture. For payment and fintech apps, the fintech security checklist covers payout and access controls. Sigi’s MVP development, QA audit and automation testing services cover this work, and the Busy Bean Coffee case study shows the testing approach in a live product.

