Vibe Coded App Audit: How to Review Your App Before Launch
AI app still broken? A developer can look at it.
Get a free assessmentFree assessment. We reply within 1 business day.
Jump to: what an audit is · before you start · step 1: built-in checks · step 2: full review prompt · the six areas · step 3: read the report · step 4: fix safely · FAQ
What a Vibe Coded App Audit Is (and What It Isn't)
A vibe coded app audit is a review of your app's code, database rules and settings, done before launch, to find the problems that would hurt you once strangers use it. Typical examples: one customer can open another customer's orders, a double click charges a card twice, a secret key sits in code the browser can download, or a page loads every row of a table at once.
AI builders write an app one prompt at a time. Each feature can work when you click through it and still leave a gap between features, such as a check that exists on one screen and is missing on the server. That is why access control matters most: in the OWASP Top 10:2025, the standard list of web app security risks, Broken Access Control is A01, the first category.
What an audit isn't:
- It isn't a penetration test. A code review reads your project. A penetration test attacks the running app from outside. Replit, for example, separates the two: its Agent security scan reviews source code, and its Level 3 scan adds a "black-box" test of the running app.
- It isn't a guarantee. Lovable's own documentation says its scanners "cannot guarantee complete security" and "do not replace a thorough security review". The same applies to any prompt, including ours.
Before You Start: Three Things to Do First
- Save a restore point. Use your builder's version history, or make a Git commit if your project is connected to GitHub. Some built-in audits change things on their own: Bolt says its audit may change your code or your database (see below). If your app already holds real user data, also back up that data with your backend's export or backup tool, because a code version is not a copy of your data. Every fix you make later is easier to undo from a known good version.
- Open the right project, on its latest version. The review only sees the project it runs in. If you have a copy or a remix, run it in the one you will publish.
- Check your credits. A full review reads your whole app, so it costs more than a normal message. Run it when you have the budget for one complete run, not halfway through a credit cycle.
What you should see: a saved version you can return to, and the project you plan to launch open in your builder.
Step 1 - Run the Security Check Your Builder Already Has
Start with the free or included checks. They are fast, they know your platform's settings, and some of them run automatically when you publish. Here is what each builder documents, checked on November 18, 2026:
| Builder | Check and where to find it | Does it change your code? | Cost |
|---|---|---|---|
| Lovable | Quick security scan runs automatically each time you publish (database access rules, the outside code packages your app installs from npm, and MCP servers, which connect AI tools to your app, left open without a login). Deep security scan reviews all application code; start it from the Security view or ask for it in the project chat. | Not by itself. Fixes happen when you apply them from Security or with "Fix all". | Scans started from the Security view are free; asking for a review in the chat uses credits. Fixes use 10 free fixes that refresh 24 hours after use, then credits. |
| Bolt | Project security audit: click Publish (top right), then Run security audit. A lighter Database security check is available on all plans. | Yes. Bolt says it "fixes most problems on its own" and may change code or the database. | Paid plans. The audit and its fixes use no tokens when started from the Publish menu; asking in chat does. Up to 30 audits a day. |
| Replit | Project Security Center: open the Tools pane, then Security Center. Replit also scans before every publish; turn on Block publishing of critical vulnerabilities to stop a release with a critical finding. | Only when you send an accepted finding to Agent. | Dependency scans are free. The Agent security scan is for paid builders; Agent fixes are billed like other Agent work. |
| Base44 | In the editor: Dashboard, then Security, then Run Security Scan. Checks permissions, private fields, exposed secrets, backend functions, dependencies and security headers. | No. Base44 says "The scan doesn't apply fixes automatically". | All plans. Static code analysis needs the Builder plan or higher. |
| Cursor | Cursor is a code editor that works on your own repository. Cursor's agent modes page doesn't describe a publish-time security scan (check Cursor's current docs if that matters to you), so the full review in Step 2 is your main check, plus the tools of your backend (below). | - | - |
If your app uses Supabase (Lovable, Bolt and Cursor projects can all use it as the backend), also open the Security Advisor in Supabase Studio. It flags database problems such as "Table publicly accessible" (a table with row level security turned off). Row level security, or RLS, is the set of database rules that decides which rows each user can read or change. Supabase's guidance is to enable RLS on every table in an exposed schema; a table without it "is readable and writable by any role with a grant on it."
What you should see: a list of findings grouped by severity, or a clean result. Fix anything marked critical before you publish, then move on to Step 2 even if the scan was clean: these checks focus on security, and most of them don't look at reliability, performance or tests.
Step 2 - Run a Full Read-Only Review With the AppStuck Master Prompt
The AppStuck master prompt is a free review prompt you paste into your builder's chat. Your AI tool then reads the whole project and writes a plain-English report in six areas, from security to testing. It is version 1.1, updated in October 2026.
- Get the prompt. Leave your email on the AppStuck master prompt page. You get it on screen straight away and a copy by email.
- Open the project you want reviewed in Lovable, Bolt, Replit, Base44 or Cursor.
- Optional extra safety net: use a mode that can't edit. The prompt already tells the AI "DO NOT modify the project. This is a read-only audit." If you want a second layer, some tools have a mode that doesn't write code: Lovable's docs say "Plan mode never modifies your code", Replit's Plan Mode (the Plan toggle at the bottom right of the chat input) "won't make any changes to your files", and Base44's Discuss mode makes "no code changes until you approve them". These modes are built for planning, so they may answer with a plan or a question instead of the report. If you don't get a report that starts with PROJECT HEALTH, run the prompt in the normal chat instead, after saving your restore point.
- Paste the whole prompt and send it. Don't shorten it. Then wait: the AI first maps your app (technologies, data flow, login, database, outside services) and only then looks for problems.
- If the AI offers to start fixing things, say no for now. Read the full report first (Step 3), so you fix the most important problem first instead of the first one it found.
How long and how much: it uses more credits than a normal message, so run it once before you launch and again after big changes, not after every small edit. In one test run of the earlier version 1.0 in Replit Agent, on one app, the review took about 9 minutes and used $0.86 of credits. Your time and cost depend on the size of your app and your plan.
What you should see: a report that starts with PROJECT HEALTH and an overall status, followed by the most important findings.
What the Review Checks: The Six Areas
These are the six areas the prompt asks your AI tool to inspect, with a few of the specific things it looks for in each. The prompt only lets the AI report a vulnerability when it can point to evidence in your project.
1. Security
Missing or inconsistent permission checks, users reaching another user's or company's data, sensitive actions protected only in the frontend, exposed keys, injection risks, unsafe file uploads, insecure webhooks (the messages services such as Stripe send to your app), payment logic that trusts the browser, and abuse risks such as unlimited expensive requests or AI calls. Why it matters: this is where data leaks and account takeovers come from.
2. Reliability and data safety
Errors that are hidden, operations that stop halfway and leave bad data, duplicate payments, webhooks, emails or records, race conditions (two actions colliding at the same moment), unsafe retries, missing timeouts, unsafe database migrations and business rules enforced only in the UI. The prompt pays special attention to what happens when something fails in the middle. Why it matters: these are the bugs that cost money and support time after launch.
3. Architecture and AI-generated code
Signs that the app has built up conflicting solutions: two login helpers, several API clients doing the same job, old and new versions of the same feature left side by side, files named old, new, final, v2 or temp, validation that differs between frontend and backend, broad error catching that hides problems, and hard-coded prices, roles or URLs. It only reports these when they create a real risk. Why it matters: each new prompt to your builder can land in the wrong copy of the code.
4. Performance and scalability
Database queries inside loops, loading whole tables, missing pagination (loading a whole list at once instead of page by page), repeated identical requests, uncontrolled background jobs, request loops and excessive AI calls or oversized prompts. Why it matters: an app that is fast with ten test rows can slow down, or run up bills, with real data.
5. Dependencies and configuration
Suspicious or apparently invented packages, conflicting packages, secrets in configuration files, inconsistent environment variable names, debug settings left on in production and risky deployment settings. The prompt tells the AI never to invent a CVE (a published vulnerability ID) without reliable evidence.
6. Testing and production readiness
Whether the behavior that matters is actually tested: login and permission failures, one user reaching another user's data, payments, destructive actions, webhooks, migrations and failures of outside services. It also checks whether the project has a credible build, test, deployment and logging setup.
Step 3 - How to Read the Report
The report has a fixed structure, so you can read it the same way every time.
| Part of the report | What it tells you | What to do with it |
|---|---|---|
| Overall status: READY, NEEDS WORK, HIGH RISK or CANNOT DETERMINE | The AI's overall view, with a short paragraph on what is most likely to cause problems. | HIGH RISK: don't launch yet. NEEDS WORK: fix the "what to fix first" list, then decide. CANNOT DETERMINE: the AI couldn't see enough; treat it as unchecked, not as safe. |
| Most important findings (10 at most) | Each one has a title, a Priority, a Confidence (CONFIRMED or LIKELY), where it is, what is wrong, why it matters and the direction of the fix. | Work top down. A CONFIRMED finding has evidence in your code; a LIKELY one needs a quick check before you spend credits on it. To check one, paste it back into the chat and ask: "Show me the exact file and lines that prove this." |
| What to fix first | 3 to 5 actions in order. | This is your launch to-do list. |
| Needs verification | Things the AI could not confirm from the code, such as production database permissions, cloud settings, backups, rate limiting or current dependency vulnerabilities. | Check each one yourself in your builder or backend dashboard, or ask a developer. Not seen is not the same as fine. |
| Positive signs | 2 to 5 things that look well built. | Leave these alone while you fix the rest. |
| Final summary | Counts per priority, the single biggest risk, and whether the AI would release today: YES, YES, WITH CONDITIONS, NO or CANNOT DETERMINE. | Use it to compare this run with the next one after your fixes. |
The four priorities, as the prompt defines them (in short):
- CRITICAL: a realistic risk of serious data exposure, account compromise, financial loss, destructive corruption, a major breach or a severe production failure. Should be fixed before release.
- HIGH: a serious problem likely to cause security, reliability, data-integrity or major scaling problems in realistic use.
- MEDIUM: a meaningful weakness likely to cause future bugs, maintenance cost, performance problems or security risk.
- LOW: a useful improvement, not a significant risk today.
What you should see: a short, ordered list of what to fix, and a separate list of things to check by hand.
Step 4 - Fix One Finding at a Time, Safely
- Start with CRITICAL, then follow "What to fix first". Don't start with the easy LOW items.
- One finding per prompt. Paste the finding's title, "Where" and "How to fix it" into your builder and ask it to fix only that. If your tool has a planning mode, ask for a plan first and read it before you approve.
- Save a version before each fix, so a bad fix costs one undo, not a day.
- Never accept a fix that makes the warning go away by weakening security. Two patterns to refuse:
- Turning off row level security, or adding a rule that lets everyone read a table, to "fix" a permissions error. Supabase: with RLS on, no data is reachable with the publishable key until you write policies for it. The fix is a correct policy, not an open table.
- Moving a secret key into frontend code. Supabase says to never use a secret key in the browser. In Lovable, anything starting with
VITE_is "a build-time, browser-exposed value"; real secrets belong in More > Settings > Secrets, which are only injected into Edge Functions (code that runs on the server) and server-side integrations.
- If a finding says a key is already exposed, moving it is not enough. Anyone who loaded your app or saw the code may have copied it. Create a new key in the service that issued it, store the new one as a secret, and revoke the old one.
- Check the fix the way a user would. For a data access finding, log in as a second test account and try to open the first account's records. Then run your builder's scan from Step 1 again.
- Run the full review again after a big batch of fixes or a big new feature, not after each small change.
What you should see: the CRITICAL and HIGH findings gone from the next run, and your builder's scan clean or down to items you have checked and accepted.
Common Mistakes When You Read a Scan or a Review
- "The builder's scan passed, so the app is safe." A clean scan means those checks found nothing. Replit says its publish-time checks "complement, rather than replace, a full Agent security scan or Level 3 scan", and Base44 says its scan "doesn't check Base44's hosting or the systems your app runs on."
- "The AI wrote it, so the AI can check it." The same model can repeat its own blind spots. That is why the prompt asks for evidence and for a separate "Needs verification" list, and why a person should look at anything involving money or personal data.
- "A checklist is enough." A manual checklist is a good second pass for security. Our guide to AI-generated app security risks covers the common flaws one by one. This guide is about the full pre-launch review, which also covers reliability, performance, dependencies and tests.
- "Read-only means free." A review doesn't change your app, but it still uses credits, because the AI reads your whole project. Plan for one full run.
What an AI Review Can't See
A code review works from what is in your project. It can't see settings that live only in a dashboard: production database permissions that were changed by hand, your hosting configuration, backups, platform rate limits or how the app behaves under real traffic. The prompt puts these under "Needs verification" instead of guessing. If your report lists several of them, or the overall status is CANNOT DETERMINE, the app hasn't been checked yet in those areas.
When to Get a Developer to Look
- The report has a CRITICAL or HIGH finding you couldn't fix in two attempts, or each fix breaks something else.
- Your app takes payments or stores personal, health or financial data.
- The status is HIGH RISK or CANNOT DETERMINE and you plan to launch soon.
AppStuck's developers fix apps built with these tools. If you want help with what the review found, see AppStuck's AI app rescue service.
What to Include When You Ask for Help
- The full report from the review, unedited. The prompt tells the AI not to show secret values, but read it once before you share it.
- Which builder you used (Lovable, Bolt, Replit, Base44 or Cursor) and which backend (Supabase, the builder's own database, or another).
- Whether the app is live, and roughly how many users it has.
- The results of your builder's own security scan from Step 1.
- What you already tried for each finding, and what changed since the review ran.
Related Guides
- Vibe coded prototype to production: the longer plan once you know what to fix.
- Lovable and Supabase errors: common database and RLS errors in Lovable apps.
- Lovable environment variables not working: where keys and secrets belong.
Frequently Asked Questions
Is my Lovable app secure?
No scan can promise that, and Lovable says its own tools cannot guarantee complete security. Start with Lovable's Quick security scan, which runs each time you publish, then run the Deep security scan from the Security view, where scans are free. Check your database access rules, and the Security Advisor if you use your own Supabase project. Then run a full read-only review of the code to cover reliability, performance and tests as well.
Is there a security checklist for vibe coded apps?
Yes. A good checklist covers who can read and change each table, keys exposed in frontend code, login and session handling, webhook signature checks, input handling and dependencies. AppStuck's guide to AI-generated app security risks has a step-by-step manual checklist. Use it together with your builder's security scan and a full code review, because a checklist only finds what it lists.
Does the AppStuck master prompt change my code?
It isn't meant to. The prompt tells your AI tool that the review is read-only and that it must not modify the project, so the tool should only read your app and write a report. An instruction is not a lock, though, so save a restore point first. For a second layer of safety, you can try a mode that doesn't write code, such as Plan mode in Lovable or Replit, or Discuss mode in Base44.
How many credits does a vibe code review use?
More than a normal message, because the AI reads your whole project. In one test run of version 1.0 of the AppStuck master prompt in Replit Agent, on one app, the review took about 9 minutes and used $0.86 of credits. Your cost depends on the size of your app and your plan, so run it once before launch and again after big changes.
What is the difference between my builder's security scan and a full review prompt?
Builder scans such as Lovable's security scans, Bolt's security audit, Replit's Security Center and Base44's security scan focus on security and on your platform's settings, and some run automatically when you publish. A full review prompt reads the whole codebase across six areas, including reliability, architecture, performance, dependencies and tests, and ranks the findings. Run both: the scan first, then the review.
Want the review for your own app? The AppStuck master prompt is free: paste it into Lovable, Bolt, Replit, Base44 or Cursor and get a plain-English list of what to fix first. Get the AppStuck master prompt.
Sources
Checked on November 18, 2026:
- Lovable docs: Security
- Lovable docs: Secrets
- Lovable docs: Plan mode
- Bolt help center: Check your project's security
- Replit docs: Project Security Center
- Replit docs: Plan vs. Build mode
- Base44 docs: Running a security scan
- Base44 docs: AI chat modes
- Cursor docs: Agent modes
- Supabase docs: Row Level Security
- Supabase docs: Database advisors
- OWASP Top 10:2025
Need Help with Your AI Project?
If you're dealing with a stuck AI-generated project, we're here to help. Get your free consultation today.
Get Free Consultation