Built your app with Claude Code, Codex, Cursor or Lovable? Citadel audits its code against OWASP and CWE: exposed secrets, vulnerable dependencies, injections, misconfigurations. You get the list of fixes to make, in priority order, then we check that they work.
Claude Code or Codex writes your app. You have no way to check what it produces.
What you get
Every flaw explained in plain words, with the exact prompt to paste into your assistant to fix it.
You can code, but you don’t know yet how to audit an app’s security.
What you get
The exact line, the OWASP and CWE reference, why it is dangerous and the expected fix.
You know the risks, but reviewing all the code before every release takes hours.
What you get
Deduplicated results, cited sources (CVE, GHSA, originating rule) and a report you can hand to your clients.
Coding with AI saves you a huge amount of time. But an assistant’s first goal is to make your app work, not to protect it. Here are the eight families of flaws it regularly produces, and that we look for in every app.
These flaws are not theoretical. Pick your type of app: each scenario starts from a real, documented flaw and shows how it turns into an incident.
They change /customers/1042 to /customers/1041 in the address bar. It works. They write a ten-line script. By morning, your full file, with phone numbers, addresses and revenue, is for sale on a forum.
The flaw
What Citadel checks
Free security check · 2 minutes
Eight questions on the flaws we find most often in apps built with AI. Your result shows up right away, no sign-up needed.
“I don’t know” counts as a risk: what nobody has checked is exactly what an attacker exploits.
Where: the file, line and code snippet involved.
Severity: scored with CVSS when the flaw is referenced, otherwise by its real impact on your app.
Confidence: how certain the detection is, so you fix the sure things first.
References: OWASP, CWE, and the CVE or GHSA ID when there is one.
Fix: the expected fix, and the prompt to paste into Claude Code, Codex or Cursor.
Verification: a new analysis confirms the flaw is gone.
Sample finding
CVSS 9.8app/api/orders/search/route.ts · lines 14 to 16 · detected by code analysis
13 const email = searchParams.get("email"); 14 const rows = await prisma.$queryRawUnsafe( 15 `SELECT * FROM orders WHERE email = '${email}'` 16 );
The risk
With ' OR '1'='1 as the email, the query returns every order of every customer. A variant can modify or drop the table.
The fix
await prisma.$queryRaw`
SELECT * FROM orders
WHERE email = ${email}`;The query becomes parameterised: user input is never interpreted as SQL again.
Fix prompt for your assistant
In app/api/orders/search/route.ts, replace prisma.$queryRawUnsafe with the parameterised prisma.$queryRaw form using a template literal. Check that no other $queryRawUnsafe or $executeRawUnsafe call receives data from the request, and add a test that sends ' OR '1'='1 as the email.
Flaws don’t hide in the same places in a Next.js, Laravel or Supabase app. Our rules target the mistakes specific to each technology, in your code (static analysis) and on your live site (dynamic analysis). 33 checks are live today.
| Check | Reference | Check type | Status |
|---|---|---|---|
| Next.js version vulnerable to middleware bypass (CVE-2025-29927) | A06 · CVE | Vulnerable dependencies | Live |
| Secret exposed in a NEXT_PUBLIC_ variable | CWE-200 · A02 | Secrets + Citadel rules | Live |
| dangerouslySetInnerHTML fed with external data | CWE-79 · A03 | Code analysis | Live |
| Redirect built from URL parameters | CWE-601 · A01 | Code analysis | Live |
| images.remotePatterns open to every domain | CWE-918 · A10 | Citadel rules | Live |
| Source maps published in production | CWE-540 · A05 | Citadel rules | Live |
| Server Actions and route handlers without a session check | CWE-862 · A01 | Citadel rules | Coming soon |
Outside the automated scope: your app’s own business logic, manual penetration testing and your hosting configuration. When your app needs them, the report tells you.
AI wrote your code: it shouldn’t be the one judging it. Detection relies on well-known open source tools and on rules hand-written by our seasoned developers.
Arthur A.
Founder of Citadel
Citadel was founded by Arthur A., a developer with years of experience, an information systems security expert and a hands-on agentic coder. He builds with AI agents every day, and has seen up close how AI can be as harmful as it is powerful.
In minutes, an agent ships a feature that works, and all too often leaves behind an admin route with no access check, a payment webhook accepted without a signature or a hand-built SQL query. The app, meanwhile, works perfectly.
What about people outside the trade? Those launching their idea with Claude Code or Codex without ever having written a line of code, those just starting out? How would they avoid these disasters?
Citadel is a collective’s answer: developers, AI engineers and cybersecurity specialists brought together by that question. Every check, every rule and every explanation is written and reviewed by the team, based on the OWASP standards. Automation runs that work on every analysis. It never decides for you.
The Citadel team
During the pilot, your first full analysis is free. The prices below will apply afterwards.
Try it on one project
$0
Vibe coders and freelancers
$9 / month
Apps in production
$25 / month
Agencies and teams
$79 / month
* Within fair use. Enterprise plans and human review: on quote.
No. Detection relies on deterministic analysis engines and our own rules: the same code gives the same result on every run. An AI can then draft a proposed fix, which you approve before anything changes.
Eight families of flaws, from access control to AI features, plus checks specific to Next.js, Node.js, Laravel and Supabase. The checks table gives, for each one, its OWASP and CWE reference, the tool that detects it and whether it is already live.
Yes. Every flaw is explained in plain words and comes with the prompt to paste into Claude Code, Codex or Cursor. A new analysis then confirms the fix works.
It is analysed in an isolated container with no internet access, then deleted when the analysis ends. Only the results are kept, with secrets masked. You can delete your projects and reports at any time.
Next.js, React, Node.js, Laravel and Supabase. Secrets and dependencies are analysed whatever the project’s language.
No. Citadel covers the flaws that can be detected automatically. For an app handling health, financial or highly sensitive data, a human audit is still recommended, and the report tells you so.
The pilot is open to a limited number of apps. Your first full analysis is free.
Rather talk first?
The Citadel letter: a common security mistake in apps built with AI, a code example, and the prompt to fix it. One email a month, unsubscribe in one click.