OWASP Top 10 for Vibe-Coded Apps

An app built with Lovable, Bolt, Cursor, Replit or v0 runs on the first try. That feels great. But “it runs” and “it is safe to launch” are two different things. Veracode’s 2025 GenAI Code Security Report found that 45% of AI-generated code introduces a known OWASP Top 10 vulnerability. And the holes are not random. They line up almost one-to-one with the OWASP Top 10 (2021). This guide maps each category to how the AI introduces it and the one-line fix, then goes deep on the three you will hit nearly every time.

Why AI code maps so cleanly to OWASP

A model trained on public code repeats the patterns it saw most. The patterns it saw most are the insecure ones. It picks the version that compiles and demos, not the version that survives an edge case. So the same mistakes land in the same places every time: permissive defaults, trusted input, secrets sitting in the client, no server-side check. Those clusters are exactly what the OWASP Top 10 catalogs. That is why a vibe-coded app fails it so reliably.

The OWASP Top 10 in vibe-coded apps

Each row takes one 2021 OWASP category, the exact way AI tools introduce it, and the shortest fix that shuts it.

OWASP category	How AI introduces it	One-line fix
A01 Broken Access Control	Scaffolds tables and APIs with no row-level security and client-side-only checks	Enforce auth and ownership on the server for every protected route and table.
A02 Cryptographic Failures	Stores passwords or tokens in plaintext, uses weak hashing, sends data over HTTP	Hash with bcrypt/argon2, encrypt sensitive data at rest, force HTTPS everywhere.
A03 Injection	Builds SQL from string concatenation and writes user input straight to the DOM	Use parameterized queries and escape or sanitize all output.
A04 Insecure Design	Skips rate limits, trusts client-supplied prices and roles, has no threat model	Validate business rules server-side and rate-limit sensitive actions.
A05 Security Misconfiguration	Leaves debug mode on, returns verbose errors, exposes admin endpoints and open CORS	Disable debug, lock down CORS, hide stack traces, close default endpoints.
A06 Vulnerable / Outdated Components	Pins old package versions with known CVEs and adds unused dependencies	Run a dependency scan, update flagged packages, remove what you do not use.
A07 Identification & Authentication Failures	Ships weak session handling, no password policy, and no brute-force protection	Use a vetted auth library, enforce strong sessions, add login rate limits.
A08 Software & Data Integrity Failures	Loads scripts from untrusted CDNs and deserializes untrusted data without checks	Pin and verify dependencies, add subresource integrity, never deserialize raw input.
A09 Security Logging & Monitoring Failures	Logs nothing useful, or logs unsanitized input enabling log injection	Log auth and security events, sanitize log inputs, alert on anomalies.
A10 Server-Side Request Forgery	Fetches user-supplied URLs server-side with no allowlist or validation	Validate and allowlist outbound URLs, block internal address ranges.

The three you will almost always hit (A01, A03, A05)

Three categories show up over and over in vibe-coded apps. Start here.

• A01 Broken Access Control. The one everyone ships: a Supabase or Postgres app where Row-Level Security was never turned on. Any logged-in user reads every other user’s rows just by changing an ID in the request. The fix is server-enforced ownership checks on every table and every endpoint.
• A03 Injection. The generated handler drops user input straight into a SQL string, or into the DOM with innerHTML. Veracode found AI failed to defend against cross-site scripting in 86% of relevant cases and against log injection in 88%. The fix is parameterized queries and escaped output.
• A05 Security Misconfiguration. Debug mode still on in production, verbose errors that hand attackers your stack traces and DB structure, CORS wide open, admin routes sitting exposed. The fix is a hardened production config that hides errors and closes the defaults.

For the deeper background on why this happens, see our pillar guide on AI-generated code security risks.

How to test for each

You do not need a security team to find most of these. Run automated scanners, add a short manual probe, and tie every finding back to its OWASP category.

• A01, A04, A10 (logic and access): manual review plus a runtime probe. Change an ID in a URL and see if you can read someone else’s data. Submit a tampered price. Point a fetch at an internal address. Tools cannot reliably catch these, so you have to poke them yourself.
• A03 Injection / XSS: a static analyzer like Semgrep for source patterns, plus OWASP ZAP for runtime XSS and injection probes.
• A02, A05 (crypto and config): OWASP ZAP and a config review for HTTPS, headers, CORS, debug flags, and error verbosity.
• A06, A08 (dependencies and integrity): Snyk or npm audit / pip-audit for known CVEs and outdated packages.
• A07, A09 (auth and logging): manual review of session handling and log output, plus TruffleHog to confirm no secrets leak into your code or logs.

For the full pre-launch sequence, see how to audit AI-generated code for security and our AI code security checklist.

Want us to run this audit for you?

We do a free 15-minute build audit. Show us your AI-built app, we tell you the exact security and production gaps and what it takes to close them. No obligation.

Book your free build audit