For founders, indie hackers and agencies shipping AI-generated apps
Built your app with AI? Hosted it? Now find out if it is actually secure.
Lovable, Bolt, v0, Replit, Cursor, Claude Code, Windsurf or Base44 wrote the code, picked the defaults and pushed it live. Nobody with a security background read a line of it. Kavach audits the live site the way an attacker would, grades it, and hands you fixes you can paste straight back into your AI tool. Nothing is attacked.
Audit my AI-built app free → See plans
Reviewed? Secure? Still in doubt?
An AI assistant optimises for "it works". It will happily open CORS to the world, leave the session
cookie unflagged, skip the Content-Security-Policy and deploy the .env file along with
everything else, because none of that stops the demo from running. These are not exotic bugs. They are
the defaults, and they are visible from outside.
What usually slips through
- The API answers any websiteGenerated backends often ship with CORS set to allow every origin so the demo 'just works'. Any page on the internet can then call your API with your users' credentials. How Kavach checks it →
- Login cookies without the safety flagsTemplates set a session cookie and stop there. Without HttpOnly, Secure and SameSite, one injected script or one http:// link is enough to steal a session. How Kavach checks it →
- No Content-Security-PolicyAlmost no AI builder emits a CSP. It is the control that stops injected scripts from running even when there is a bug in the code. How Kavach checks it →
- The .env file went live with the buildThe assistant put secrets in .env, the deploy copied the whole folder, and the keys are now one URL away. Kavach checks this for verified owners without ever storing what it sees. How Kavach checks it →
- The .git folder is on the web serverFull source history, published. Anyone can download the code the AI wrote, including every mistake and every secret ever committed. How Kavach checks it →
- Headers advertise the framework versionHosting defaults announce the exact server and framework build, which tells an attacker precisely which known exploits to try. How Kavach checks it →
- Your domain can be spoofedThe app sends email, but nobody set SPF, DKIM or DMARC. Phishing mail 'from' your domain lands in your customers' inboxes and looks real. How Kavach checks it →
- HTTPS is on, but not enforcedThe hosting gave you a certificate. It did not tell browsers to refuse plain HTTP, so a coffee-shop attacker can still downgrade the first request. How Kavach checks it →
How it works
- Paste the live URLNo install, no code access, no agent. Kavach only looks at what the public internet can already see.
- Get a letter grade in secondsEvery finding comes with the plain-English attack it enables and a severity, weighted for AI-built apps.
- Paste the fix back into your AI toolThe Fix Pack is copy-paste config for your stack. Give it to the same assistant that built the app and re-scan.
- Keep it that wayVerify the domain, schedule re-scans, and get an alert the moment a new prompt or deploy opens something up.
Also scan the code itself
The live-site audit sees what an attacker sees. Code Scan looks at what the AI wrote: zip the project, upload it, and Kavach streams through every file for string-built SQL, shell commands with user input, unescaped output, credentials in source, wide-open CORS, disabled TLS checks, unsafe deserialisation, debug left on and platform rules (Firebase, Supabase) that let everyone in. Any language, hundreds of thousands of lines, nothing extracted to disk, and the archive is deleted the moment the scan ends. You get a graded report and a remediation prompt to paste into the same AI tool. On any finding you can also Ask Kavach AI: a second opinion (likely real or likely false positive) and a draft of the fixed line. Only that one excerpt is processed, by a model on our own server or by an inference provider that does not train on your data; the answer says which.
What Kavach does not do
Kavach does not log in, does not test business logic and does not submit data to your forms. The live-site audit is a non-destructive, external check of the controls that stop the most common attacks on AI-built apps; Code Scan is pattern-level static analysis, not a line-by-line human review. A good grade is necessary, not sufficient. Results are observations, not certifications or guarantees; see the disclaimer.
Questions founders ask
Does Kavach review my code?
Two ways. The live-site audit never needs your repository, hosting login or API keys, so it works even for platforms that do not hand you the code. Code Scan takes a zip of the project and runs pattern-level static analysis across every file; the archive is deleted as soon as the scan finishes and only the findings are kept.
Is the scan safe to run on my live app?
Yes. Every check is read-only: it fetches pages and headers, inspects TLS and DNS, and looks for files that should not be public. Nothing is submitted, brute-forced, fuzzed or taken down. Checks that could reveal secrets (.env, .git) only run after you prove you own the domain, and Kavach never stores what it finds there.
I do not understand security terms. Will the report make sense?
That is the point. Each finding says what it is, why it matters, what an attacker would do with it, and the exact text to paste into your config or back into your AI assistant.
Will an A+ mean my app is secure?
No tool can promise that, and Kavach does not. An A+ means the externally visible controls are in place. Authorisation bugs, business-logic flaws and anything behind a login are outside what a non-destructive external scan can see. Treat the grade as the first gate, not the last word.
Can I use it on an app I built for a client?
Yes, if you are authorised to test it. Agencies get client rooms, white-label PDF reports and a public Trust Passport page they can hand to the client.