OWASP Top 10 for vibe coding: the 2025 list, read for an AI-built app

The OWASP Top 10:2025 is the list to use for a vibe-coded app, and OWASP names vibe coding on the list's Next Steps page as X03:2025, Inappropriate Trust in AI Generated Code. In apps built by prompting, three categories carry most leaks: broken access control (A01), security misconfiguration (A02) and authentication failures (A07). Half of the list can be checked on your live app in an hour.

The team behind Keybrum's attacksUpdated 14 min read

In short

  • The OWASP Top 10:2025 keeps Broken Access Control first, moves Security Misconfiguration from fifth to second, and adds two categories: Software Supply Chain Failures and Mishandling of Exceptional Conditions.
  • OWASP names vibe coding itself in X03:2025, outside the ten. Its advice starts with one line: you are responsible for all code that you commit.
  • Tests of coding agents match the list. In January 2026 Tenzai found no exploitable SQL injection or XSS in 15 apps built by five agents, and authorization flaws in many of them.
  • A01, A02, A05 and A07 show on the deployed app, so you can test them from outside. A03, A08 and A09 live in your lockfile, your pipeline and your logs.
  • A scanner that reports no OWASP issue has not signed in as a second user. Access control between accounts is the test that needs two identities.
On this page

Is there an OWASP Top 10 for vibe coding?

The list to use is the OWASP Top 10:2025, the eighth edition of the ranking of web application risks. OWASP built it from data on more than 2.8 million applications, and it applies to an app whatever wrote the code. We found no separate Top 10 for vibe coding on OWASP's sites in November 2026. What OWASP did is name vibe coding inside the 2025 edition, on a page called Next Steps that lists three risks that did not make the ten.

The third of them is X03:2025, Inappropriate Trust in AI Generated Code ('Vibe Coding'). OWASP describes code written and committed almost entirely without human oversight, and notes that AI generated code often contains more vulnerabilities than code written by people. Its first recommendation is short.

You are responsible for all code that you commit.
OWASP Top 10:2025, X03 Inappropriate Trust in AI Generated Code

OWASP publishes other lists, and each answers a different question. Pick by what you are worried about.

Your questionThe OWASP documentWhere to read more
Can a stranger read or change data in my app?OWASP Top 10:2025This page
My app calls a model. What can go wrong with the prompts?OWASP Top 10 for LLM ApplicationsThe LLM Top 10 as questions about your app
My agent has tools and access. What can it be made to do?OWASP Top 10 for Agentic ApplicationsAI agent security
Is my coding session itself safe?Secure Coding with AI Cheat SheetSlopsquatting, for the dependency part

The OWASP Top 10:2025, category by category, in an app built by prompting

The 2025 edition has two new categories and one merge. Software Supply Chain Failures replaces the older category on outdated components and widens it. Mishandling of Exceptional Conditions is new. Server-side request forgery, a category of its own in 2021, now sits inside Broken Access Control. Here is the list, with the form each risk takes in an app built with Lovable, Bolt, Replit, v0, Cursor or Claude Code.

CategoryWhat it looks like in a vibe-coded appVisible from outside?
A01 Broken Access ControlA table without row level security, a policy that checks sign-in but not ownership, a route that returns record 42 to whoever asksYes
A02 Security MisconfigurationA served .env file, public source maps, a debug route left on, missing security headers, a public storage bucketYes
A03 Software Supply Chain FailuresAn old Next.js or React version with a published flaw, a package the model inventedPartly: a known flaw can be proven on the live app
A04 Cryptographic FailuresA default or guessable signing secret, an old TLS version still accepted, no HSTSPartly
A05 InjectionRaw SQL built from input in a database function, HTML rendered without escapingYes
A06 Insecure DesignA checkout that accepts a negative quantity, a client that sends its own price or rolePartly
A07 Authentication FailuresA secret key in the JavaScript, a login checked in the browser, no limit on sign-in attemptsYes
A08 Software or Data Integrity FailuresA payment webhook that never verifies the signature, a script from a CDN without an integrity hashYes
A09 Security Logging and Alerting FailuresNo record of failed sign-ins, no alert when your AI bill triples overnightNo
A10 Mishandling of Exceptional ConditionsAn API error that prints a stack trace, an auth check that lets the request through when it throwsPartly
Category names are OWASP's. The examples are ours, drawn from the holes that recur in AI-built apps.

The three categories that carry most leaks

A01 Broken Access Control: who may read which row

OWASP keeps this category first and calls it the most serious application security risk. Its description reads like a list of what agents skip: access open to anyone instead of denied by default, viewing someone else's account by supplying its identifier, an API with no access control on POST, PUT and DELETE, and reaching privileged pages by guessing the URL.

In an AI-built app the usual form is a database the browser talks to directly. On Supabase, what stands between a visitor and your tables is row level security (RLS), the rules that decide which rows each caller may see. The scan behind CVE-2025-48757 found 303 unprotected endpoints across 170 of 1,645 Lovable showcase apps, about one in ten. Tests of the agents point the same way. In January 2026 the security firm Tenzai had five coding agents, Cursor, Claude Code, Codex, Replit and Devin, build the same three apps from the same prompts. It counted 69 vulnerabilities in the 15 apps. Improper authorization on API calls was among the most common, and all five agents introduced a server-side request forgery flaw in the test built for it.

Check your own project. On Supabase, run this in the SQL editor. It reads the catalog and changes nothing:

sql
-- 1. Tables in public with row level security off
select c.relname as table_name
from pg_class c
join pg_namespace n on n.oid = c.relnamespace
where n.nspname = 'public'
  and c.relkind in ('r', 'p')
  and not c.relrowsecurity
order by 1;

-- 2. Policies that let every caller through
select tablename, policyname, cmd, roles, qual, with_check
from pg_policies
where schemaname = 'public'
  and (qual = 'true' or with_check = 'true');
Every row of the first result is a table anyone with your public key can reach. Every row of the second is a policy that filters nothing.

Then run the test no query replaces: create two accounts, sign in as the second, and try to open the first one's records, in the app and by calling the API with the second account's session. The Supabase RLS guide has the policies for the common cases, the RLS checker reviews the ones you paste, and the IDOR guide covers routes that take an id.

A02 Security Misconfiguration: what the host answers

This category moved from fifth place to second. OWASP's explanation is that more of an application's behaviour now depends on configuration. That describes a generated app well: the database rules, the storage buckets, the sign-up settings and the hosting options are all settings, and most of them are not in the repository. OWASP's own list includes unnecessary features left enabled, error messages that reveal stack traces, and missing security headers. Its prevention advice also says to avoid static keys or secrets in code and config files.

Three requests against your own domain show the basics:

bash
# Each should answer 404, or the same page as a path that does not exist
curl -s -o /dev/null -w "%{http_code}\n" https://your-app.example/.env
curl -s -o /dev/null -w "%{http_code}\n" https://your-app.example/.git/HEAD

# The security headers your app sends
curl -sI https://your-app.example | grep -iE "strict-transport|content-security|x-frame|x-content-type|referrer-policy"
Run these against your own app only. Single-page apps often answer 200 with the home page for any path: compare the body, not only the status.

The exposed .env guide covers the fix per host, and the security headers guide explains each header.

A07 Authentication Failures: keys and logins

OWASP maps hard-coded credentials to this category, next to weak passwords, sign-in attempts that are never slowed down, and sessions that stay valid after logout. In a generated app the first one is the classic: the feature needed an OpenAI key or a Supabase service role key, and the shortest path to a working screen was to put it in the browser. Wiz Research, studying vibe-coded apps in September 2025, also found logins checked entirely in the browser, with the password readable in the JavaScript.

Search your built JavaScript for sk-, sk_live_, service_role and -----BEGIN, or paste it into the API key leak checker. A key that shipped in a public bundle is burned: rotate it with the provider, then move the call behind a server route. The guide on API keys exposed in the frontend lists which keys are safe to ship, and the rate limiting guide covers the sign-in route.

The categories your agent handles better, and the catch

A05 Injection is where generated code does best. Tenzai found no exploitable SQL injection or cross-site scripting in its 15 apps. Frameworks do much of that work: React escapes what it renders, and database clients send values as parameters. The catch shows when the framework is not in the way. Veracode's Spring 2026 update, which tests short coding tasks without security instructions, reports that models wrote safe code in 82% of SQL injection tasks and only 15% of cross-site scripting tasks. In your app, look at the places that leave the safe path: a database function that builds SQL from text, and any use of dangerouslySetInnerHTML. For features that call a model, OWASP's injection page points to a separate risk, prompt injection.

A04 Cryptographic Failures is mostly handled by your platform: the host serves HTTPS, and the auth service hashes passwords. Veracode's models picked a safe algorithm in 86% of cryptography tasks. What remains is yours: a signing secret left at a default value on a self-hosted backend, a private key committed to the repository, and the HSTS header that tells browsers to stay on HTTPS.

The categories nobody prompts for

A03 Software Supply Chain Failures. Half of the respondents to OWASP's community survey ranked this category first. For a generated app it has two forms. The first is an old framework version: the project starts from whatever the template or the agent picked, and nothing in the prompting loop upgrades it. CVE-2025-55182, a remote code execution flaw in React Server Components, is the example to remember. The second is a package that does not exist until an attacker registers it, which the slopsquatting guide covers. Run npm audit and npm ls next react, and compare with the Next.js CVE checker.

A06 Insecure Design. OWASP's line is that an insecure design cannot be fixed by a perfect implementation. An agent implements what the prompt describes, and prompts rarely describe abuse. In Tenzai's test, four of five agents accepted an order with a negative quantity and three of five accepted a product with a negative price, because nothing in the prompt forbade it. Try both on your own checkout in test mode, and check that the server recomputes the price instead of trusting the one the browser sent.

A08 Software or Data Integrity Failures. The common case is a payment webhook that parses the JSON and marks the order paid without verifying the signature. Anyone who finds the URL can then send a fake event. The Stripe webhook guide has the handler to paste.

A09 Security Logging and Alerting Failures. OWASP renamed this category to stress alerting: a log nobody reads does not help. A generated app usually has neither. You will not see this from outside, and neither will a tester. The minimum is a spend alert on every paid API key, your database and host logs kept long enough to answer "since when", and a short plan for the day something leaks.

A10 Mishandling of Exceptional Conditions. The new category covers error messages that leak details and code that fails open, meaning a failed check leaves the door open instead of closed. The generated form is a try block around the session check whose catch logs the error and carries on. Send your own API route a body it cannot parse and read what comes back:

bash
curl -s -X POST https://your-app.example/api/your-route \
  -H "content-type: application/json" \
  -d '{"broken"'
The right answer is a short error. A stack trace, a file path or a table name tells a stranger how your server is built.

X03:2025: what OWASP says about vibe coding itself

X03 has no list of weaknesses attached. When OWASP wrote it, it noted that no CVE or CWE related to AI generated code existed yet. That has started to change: Georgia Tech's Vibe Security Radar, which traces published vulnerabilities back to commits signed by an AI tool, had confirmed 74 CVEs tied to AI coding tools by late March 2026, with 35 counted in March alone, according to Infosecurity Magazine.

OWASP's recommendations are written for developers. Be able to read and fully understand all code you submit. Review AI-assisted code yourself and with static analysis. Keep a list of prompts that take security practices into account: the guide to secure vibe coding has a rules file and three prompts to start from. OWASP adds that vibe coding is not recommended for complex functions, business critical programs, or programs that are used for a long time.

A one-hour OWASP pass on your own app

  • A01: the RLS query returns no table, and the policy query returns no policy on a table that holds user data.
  • A01: signed in as a second account, you cannot open the first account's records, in the app or through the API.
  • A02: /.env and /.git/HEAD on your domain answer like a page that does not exist, and no .map file sits next to your JavaScript.
  • A03: npm audit shows no critical advisory, and your Next.js and React versions are patched.
  • A05: no database function builds SQL from text, and no user content goes through dangerouslySetInnerHTML.
  • A06: your checkout refuses a negative quantity, and the server sets the price.
  • A07: your built JavaScript holds no secret key, and ten wrong passwords in a row get slowed down or blocked.
  • A08: your payment webhook answers 400 to an event without a signature.
  • A09: a spend alert exists on every paid API key, and you know where your logs are.
  • A10: a malformed request to your API returns a short error with no stack trace.

The vibe coding security checklist goes through the same ground layer by layer, with the fix for each item.

What Keybrum tests from this list, and what it does not

Keybrum attacks the deployed app, signed out and then as accounts it creates. It covers the categories that show from outside, and it does not see the ones that live in your repository or your logs.

CategoryWhat Keybrum attacks on the live appNot covered
A01Reads, inserts and updates on Supabase tables with the public key, as a stranger and as a new account, with every write rolled back. Database functions, storage buckets, Firebase rules. One account reading or changing another account's object. Admin pages opened signed out. The Next.js middleware bypass. Server-side request forgery, proven by a callbackBusiness rules only you know, such as which team may see which project
A02.env files, the .git folder, source maps, backup and database files, debug routes, the security header baseline, CORS that accepts any origin with credentialsSettings that never answer on the internet
A03Next.js and React flaws, proven by how the app behaves, never guessed from a version numberYour lockfile, invented packages, your build pipeline
A04A weak or default JWT secret on Supabase, a guessable session signing secret, the certificate and old TLS versions, HSTSEncryption at rest, password hashing
A05SQL injection confirmed by a database error only raw SQL produces, HTML injection reflected and stored, template injection, a sign-in that accepts a query operator in place of a passwordInjection in code paths the public app never exposes
A06Negative and zero values on payment fields, a client that sets its own role, plan, price or credits, one-time actions accepted several times at onceA review of your design
A07Secret keys in the bundle and in API answers, checked with read-only calls. Sign-in without throttling, answers that reveal which emails are registered, a session still valid after logout, tokens accepted with no valid signatureMulti-factor setup, your identity provider's own settings
A08Stripe, Polar, Lemon Squeezy, Paddle, GitHub and Clerk webhooks that accept an unsigned event. CDN scripts without an integrity hashYour CI pipeline and update process
A09Nothing: logging and alerting cannot be seen from outsideAll of it
A10API errors that print a stack trace or name tables, columns and server filesFail-open logic in general

Every finding comes with the request that worked and a fix written for your coding agent, and Keybrum attacks again after the fix. It only attacks apps whose owner proved ownership. A one-time report costs €49. Plans start at €29 a month excluding VAT and attack again on every deploy. Keybrum's code audit, which includes Cloudflare's security-audit skill, reads the repository for the causes. Neither one reviews your design for you, and a clean run means those attacks failed on that date, nothing more.

Questions

Is there an official OWASP Top 10 for vibe coding?

We found none on OWASP's sites in November 2026. The OWASP Top 10:2025 applies to any web app, and its Next Steps page names vibe coding as X03:2025, Inappropriate Trust in AI Generated Code. OWASP also publishes a Secure Coding with AI cheat sheet about the coding session itself.

What is X03:2025 in the OWASP Top 10?

It is one of three risks listed after the ten, on the Next Steps page of the 2025 edition. It covers code written and committed almost entirely without human oversight. OWASP's advice is to read and understand all code you submit, to review it with your own eyes and with static analysis, and not to use vibe coding for complex or business critical programs.

Which OWASP Top 10 risk is most common in vibe-coded apps?

Broken Access Control, A01. The public cases come down to it: tables readable with the public key, and users able to open each other's records. In Tenzai's January 2026 test of five coding agents, improper authorization on API calls was among the most common flaws.

Does AI-generated code still have SQL injection?

Less than other flaws. Tenzai found no exploitable SQL injection or XSS in 15 apps built by five agents, and Veracode's Spring 2026 tests show models writing safe code in 82% of SQL injection tasks. Check the places that leave the framework's safe path, such as a database function that builds SQL from text.

What changed in the OWASP Top 10 in 2025?

Two categories are new: Software Supply Chain Failures (A03) and Mishandling of Exceptional Conditions (A10). Security Misconfiguration moved from fifth to second. Server-side request forgery was folded into Broken Access Control, which stays first.

Is my app safe if a scanner reports no OWASP Top 10 issue?

It means the checks that scanner ran found nothing. Most scanners do not sign in as two users, so they cannot see one account reading another's data, and none can see your logging. Ask what was tested, and run the two-account test yourself.

Sources

  1. 1.OWASP Top 10:2025, OWASP
  2. 2.OWASP Top 10:2025, Introduction: what changed in 2025, OWASP
  3. 3.Next Steps: X03:2025 Inappropriate Trust in AI Generated Code ('Vibe Coding'), OWASP
  4. 4.A01:2025 Broken Access Control, OWASP
  5. 5.A02:2025 Security Misconfiguration, OWASP
  6. 6.A07:2025 Authentication Failures, OWASP
  7. 7.A10:2025 Mishandling of Exceptional Conditions, OWASP
  8. 8.Secure Coding with AI Cheat Sheet, OWASP Cheat Sheet Series
  9. 9.Vibe Coding Tested: AI Agents Nail SQLi but Fail Miserably on Security Controls, SecurityWeek
  10. 10.Spring 2026 GenAI Code Security Update, Veracode
  11. 11.Security Researchers Sound the Alarm on Vulnerabilities in AI-Generated Code, Infosecurity Magazine
  12. 12.Wiz Research discovers one in five organizations exposed to systemic risks in vibe-coded applications, Wiz Research
  13. 13.Statement on CVE-2025-48757, Matt Palmer
  14. 14.CVE-2025-48757, NIST National Vulnerability Database

Your app is public. So are its mistakes.

Check what any attacker can already see in it, free, in ten seconds. The full attack, with proof and fixes: €49 once, or €29 a month on a plan.

or continue with Google or GitHub