Reference · OWASP Top 10:2025

The OWASP Top 10, explained for code that AI wrote

The ten most critical web application risks, in the 2025 edition, one page each. Every page covers what the risk is, how it shows up when the code came from Cursor, Claude, Copilot, Lovable or Bolt, a before-and-after example, how to find it, how to fix it, and which fix guides apply. Where we have data, from 2,148 scanned repos and 4,740 probed apps, it is on the page.

What changed in 2025

Two new categories: Software Supply Chain Failures at A03 and Mishandling of Exceptional Conditions at A10. SSRF folded into Broken Access Control. Security Misconfiguration up to A02. Logging and Monitoring became Logging and Alerting. Broken Access Control stays at number one, and in AI-built apps it is not close.

A01Broken Access ControlUsers can act outside their intended permissions: read other people's data, reach admin pages, call endpoints they should not.→A02Security MisconfigurationThe app is built right but deployed wrong: debug on in production, default credentials, permissive CORS, missing headers, files the server should not serve.→A03Software Supply Chain FailuresThe code you did not write: vulnerable or malicious dependencies, compromised build pipelines, and packages that do not exist until an attacker publishes them.→A04Cryptographic FailuresData that should be protected is not: weak or missing encryption, secrets hardcoded in source, passwords hashed badly, tokens generated with Math.random().→A05InjectionUntrusted input reaches an interpreter as code: SQL, NoSQL, shell commands, eval, HTML, templates, LDAP.→A06Insecure DesignThe code works as designed and the design is the problem: no rate limit on login, a password reset that trusts the email in the request, business logic with no abuse case.→A07Authentication FailuresProving who you are, done badly: JWTs decoded but not verified, sessions in localStorage, no brute-force protection, weak password rules, tokens that never expire.→A08Software or Data Integrity FailuresTrusting code or data without checking it came from where you think: unverified webhooks, unsafe deserialization, CI that pulls unpinned steps, auto-updates without signatures.→A09Security Logging and Alerting FailuresYou cannot see the attack: logins and failures not logged, logs that contain secrets, no alert when something goes wrong, or no logs at all.→A10Mishandling of Exceptional ConditionsWhat the app does when something goes wrong: fails open, leaks a stack trace, catches and ignores, or crashes on input it did not expect.→

Where AI-built apps actually fail

The list is ranked by OWASP's data across all applications. Our data on AI-built apps ranks differently. Hardcoded credentials (A04) were the top high-severity finding in every group of repos we scanned, 276 of 2,148. Broken access control (A01) is the top live finding: 261 of 4,740 deployed apps served a private page, API records or a readable database to a logged-out visitor. Injection (A05) follows, then misconfiguration (A02) as served environment files and missing headers.

The reason is consistent across categories: AI tools optimise for code that works, insecure code works, and the prompt rarely mentions the abuse case. The category pages are written around that.

Frequently asked questions

What is the OWASP Top 10?

A list of the ten most critical web application security risks, published by the Open Worldwide Application Security Project and updated every few years from data on real tested applications. It is the reference most security standards, audits and scanners map to. The current edition is OWASP Top 10:2025.

What changed in the 2025 OWASP Top 10?

Two new categories and one merge. Software Supply Chain Failures (A03) and Mishandling of Exceptional Conditions (A10) are new. Server-side request forgery, which had its own slot in 2021, is now part of Broken Access Control. Security Misconfiguration moved up to second, and Logging and Monitoring was renamed to Logging and Alerting.

Which OWASP categories matter most for AI-generated code?

From our scans, three dominate: Cryptographic Failures, because hardcoded credentials are the top finding in every group of AI-built repos; Broken Access Control, because AI tools build the login and skip the authorization check, which is why 1 in 18 live AI-built apps served a private page to a logged-out visitor; and Injection, in the corners where the model falls back to string concatenation. Security Misconfiguration follows, mostly as served .env files and missing headers.

Does Prbl cover the whole OWASP Top 10?

It covers the parts a scanner can: A01, A02, A04, A05, A07 and A08 directly through rules and live probes, A03 for hallucinated packages, and the implementation traces of A06, A09 and A10 (missing rate limiting, sensitive data in logs, leaked stack traces, ReDoS). Design review and alerting are human work; each category page includes a checklist for that part.

Is the OWASP Top 10 a compliance standard?

Not formally, but it is what most standards point at. SOC 2 auditors, PCI DSS, and customer security questionnaires routinely ask how you address the OWASP Top 10, so being able to say which controls cover which category is useful evidence.

Check your own app against the categories a scanner can see. Live URL or public repo, free, no account.

Scan my app →

Related: What is the OWASP Top 10? (short version) · SAST for AI code · all 53 fix guides · Next.js security checklist

OWASP Top 10 (2025) Explained for AI-Generated Code: Every Category, With Examples and Fixes | Prbl