Skip to main content

Security Audit for a Vibe-Coded App: What We Check

By the OnlineAppCreator editorial team4 min read

Short answer

A security audit of a vibe-coded app focuses on the weaknesses code generators introduce most: broken access control, exposed secrets, client-side-only checks, unsafe database policies, vulnerable dependencies and unprotected AI endpoints. We combine automated scanning with manual testing as different users, review infrastructure and deliver prioritised findings with fixes that agents can implement and humans verify.

On this page
  1. Why vibe-coded apps have a distinct risk profile
  2. Scope: what we check
  3. How the audit runs
  4. The findings we see most often
  5. After the audit

Vibe coding, building software mainly by prompting an AI and accepting what works, has made it possible for non-engineers to launch real apps. It has also produced a wave of apps with avoidable security holes: data readable by any visitor, admin functions callable by anyone, API keys in the browser. Most of these are not sophisticated vulnerabilities. They are basic controls the generator never added and nobody checked. This article describes what a security audit of a vibe-coded app should cover, based on the checks we run before taking an AI-built product into operation.

Why vibe-coded apps have a distinct risk profile

AI generators optimise for a working result. When an error appears, the fastest fix is often to loosen a permission, move a check to the frontend or disable a validation. The app keeps working, so the change is accepted. Over dozens of iterations, security decisions are made implicitly and undocumented. In addition, the person directing the build may not know which parts of the code are security relevant. The audit's job is to make these implicit decisions visible and to correct them.

Scope: what we check

Access control and authorisation

This is the most common and most severe category, consistent with Broken Access Control ranking first in the OWASP Top 10. We test whether user A can read or change user B's data, whether normal users can reach admin functions, whether IDs in URLs can be changed to access other records and whether checks exist on the server, not just in the interface. For Supabase-based apps, this includes every Row Level Security policy and storage bucket; our Supabase RLS guide covers that in depth.

Authentication and sessions

Password reset flows, email verification, token lifetimes, logout behaviour, rate limiting on login and sign-up, and multi-factor options for admin accounts.

Secrets and configuration

Search of the frontend bundle, repository history and environment variables for keys that should not be public, such as service role keys, payment secrets or AI API keys. Keys that were ever exposed are rotated, not just removed.

Input handling and injection

SQL built from strings, unsanitised HTML rendering that enables cross-site scripting, file uploads without type and size checks, server-side requests to user-supplied URLs and open redirects.

Dependencies and supply chain

Known vulnerabilities in packages, abandoned or typo-squatted packages and packages an agent suggested that do not exist in the expected form. Lockfiles must be committed and audited in CI.

AI features

Endpoints that call language models need authentication, rate limits and spend caps. We test for prompt injection where model output triggers actions or is rendered as HTML, and check that system prompts and tool permissions do not expose data across users.

Infrastructure

Security headers, TLS configuration, CORS rules, separation of staging and production, backups, logging of security events and who has access to dashboards and hosting accounts.

How the audit runs

  1. Kick-off: architecture walkthrough, list of user roles, data categories and external services.
  2. Automated pass: static analysis, dependency audit, secret scanning, header and TLS checks.
  3. Manual testing with multiple test accounts per role against every API endpoint and database table.
  4. Code review of authentication, authorisation, payments and AI integrations, assisted by agents to map the codebase quickly.
  5. Report with findings ranked by severity and exploitability, each with a concrete fix and a test that proves it.
  6. Re-test after fixes and a short summary suitable for customers or investors.

Agents make the audit faster but not optional. Claude Code or Codex can map every endpoint, list every database query and draft fixes. The judgement about what an attacker can actually do still requires an experienced reviewer.

The findings we see most often

In vibe-coded apps the pattern is consistent: tables or buckets readable by any authenticated user, admin flags editable through the profile update endpoint, payment status trusted from the client, AI endpoints without rate limits, secrets in frontend environment variables and no separation between test and production data. Less frequent but serious are debug endpoints left enabled, verbose error messages that reveal stack traces and database structure, and webhooks that accept requests without verifying the sender's signature. Almost all of them are cheap to fix once found. The expensive part is discovering them after a customer or researcher does.

After the audit

A one-time audit is a snapshot. Keep the improvements by adding the tests from the report to CI, documenting security rules in the agent context file and repeating a lighter review before major releases. Teams that want to build these habits can train them in the neooutput secure agentic coding workshops. Security also connects to GDPR obligations, covered in our GDPR guide for AI-built apps, and to the broader hardening path in from AI prototype to production.

If you want an independent audit or need the fixes implemented, the OnlineAppCreator custom development team performs security reviews of AI-built apps under ISO 9001 certified processes and supports remediation and ongoing operation, as part of how we turn ideas into operating products.

Frequently asked questions

What is the most common vulnerability in vibe-coded apps?

Broken access control: users can read or modify other users' data, or reach admin functions, because checks are missing on the server or in database policies.

Can an AI agent do the security audit itself?

Agents speed up mapping and scanning and can draft fixes, but they miss business-logic flaws. An experienced human reviewer should test as an attacker and verify results.

How long does a security audit of a small app take?

For a typical small app with a few roles, an audit including manual testing and a report usually takes a few days. Larger apps with many integrations take longer.

What should I do if a secret key was exposed?

Rotate it immediately at the provider, remove it from code and history where possible, check logs for misuse and move it to a server-side secret store.

Want this done right, the first time?

OnlineAppCreator takes products from idea to operation. Tell us where you stand and get a concrete plan in a free 15-minute call.