Cybersecurity Audit for a Vibe-Coded App: Do It at Launch
AI-built apps ship with known security holes, and attacks on Australian businesses keep coming. Why a security review belongs before or right after launch.
On Monday 16 November, Coles told the share market it was dealing with a cyber incident. Two weeks earlier, NSW Police confirmed its Cybercrime Squad was investigating a breach at NSW Health. Both have security teams, budgets and people whose whole job is this.
Now think about the app you built with Lovable, Bolt, Replit, v0 or Cursor. Nobody has reviewed what the AI generated, and it is about to hold real customers' names, emails and payments, or it already does.
This post is about that gap: what the research says about security holes in vibe-coded apps, why the current run of attacks makes it urgent, and why a security review is the first step of a vibe code cleanup, best done before launch or straight after.
Short answer: Vibe-coded apps ship with predictable security holes: database tables anyone can read, secret keys in the browser, permissions checked only on screen. Independent scans keep finding them in most apps tested. A cybersecurity audit of the app (a review of the code and data by someone who reads code) finds them in days. The right time is before real customer data goes in, or right after launch if it already has.
How many holes are we talking about?
You do not have to take a developer's word for it. Three pieces of research from this year, with their sources:
A scan of 1,072 live vibe-coded apps. Symbiotic Security scanned apps built with Lovable, v0, Bolt, Replit and Windsurf on Supabase and reported that 98% had at least one security flaw and 16% had a critical one. On 172 of them, anyone could delete records without signing in. On 39, every table was readable by anyone. (Symbiotic Security, 2 June 2026.)
The Lovable vulnerability. A Cloud Security Alliance research note describes CVE-2025-48757: a scan of 1,645 Lovable apps found 303 vulnerable endpoints across 170 applications, because the database rules that stop strangers reading tables were never switched on. The note says Lovable "did not implement RLS unless explicitly instructed", and founders did not know to ask. (Cloud Security Alliance, 6 April 2026.)
AI-generated code in general. The same note cites Veracode's 2025 GenAI Code Security Report, which found that 45% of AI-generated code samples failed security tests against the standard list of web application risks.
A fair caveat: the first scan is published by a company that sells security tools, and its sample was apps it could find, not a random draw. Treat the percentages as indicative. The pattern is the point, and every one of these sources describes the same handful of holes.
Why AI builders leave them
An AI builder optimises for one thing: making the screen you described appear. An open database never breaks a screen, so nothing ever prompts the tool, or you, to close it. The holes we find most often on a cleanup are the same ones the research lists:
- Permissions checked in the browser. The app hides the admin button, but the server never checks who is asking. Anyone who opens developer tools can read or change other people's data.
- Database rules missing or wide open. Row level security in Supabase, or its equivalent elsewhere, switched off or written to let everything through.
- Keys and secrets in the frontend. API keys shipped to every visitor's browser, where they can be copied and billed to you.
- No backups and no monitoring. If the database is wiped or the site falls over at night, nobody is told and there is nothing to restore.
None of this means the tool is bad or you were careless. A prototype's job is to prove an idea. The moment real people's data goes in, the app has a different job.
Why now: the attacks are not slowing down
The Australian figures published in the last month all point the same way.
- The ASD's Annual Cyber Threat Report 2025-26 counts more than 82,000 cybercrime reports in the year, about one every six minutes. The average self-reported cost per report was $55,400 for a small business and $141,300 for a medium business, up 45%. (business.gov.au, 28 October 2026.)
- The privacy regulator's breach figures for January to June 2026 show 667 notifications, with malicious or criminal attacks the leading cause, and breaches affecting more than a million people up from four to 11. (techpartner.news, 29 October 2026, reporting the OAIC's statistics.)
- NSW Health, 30 October: a breach "impacting an isolated server", declared a major incident, with officials saying no patient data had been impacted at that time. (SBS News, 4 November 2026.)
- Coles, 16 November: a cyber incident disclosed to the ASX. Stores and online kept running, and Coles said it had found no evidence so far that customer, team or supplier data was compromised. (iTnews, 16 November 2026.)
Look at why those two incidents stayed contained. The server was isolated. Operations were separate from whatever was hit. Those are decisions made long before the attacker turned up.
A vibe-coded app has usually had none of those decisions made. And attackers do not pick targets by fame. The scans above were run by researchers with automated tools; the same tools, pointed at the same public addresses, are what an attacker uses. A new app with an open database does not need to be noticed by a person to be found.
The first step of a vibe code cleanup is the security check
A vibe code cleanup starts with a prototype review, and security is the first thing we look at in it. We read the code and the data, and before anything else we answer four questions:
- Can a logged-in user reach another user's records?
- Can someone with no account read or change anything at all?
- Are any secret keys visible in the browser?
- Is there a backup, and has anyone restored from it?
You get the answers in writing, with a fixed price and timeline for fixing what we found, before you commit to anything. Sometimes the verdict is a focused pass on permissions, keys, backups and monitoring. Sometimes the data model or the permissions are the problem and the foundation needs rebuilding underneath, while the screens your users know stay the same and the prototype keeps running. We will say which, and say so if the foundation is sound.
This is a cybersecurity audit of your app, not of your whole business. We are a software studio, not a managed security provider: it is a review of the code and configuration, not a penetration test or a certification, and your laptops, email and network still belong with an IT or security firm.
Before launch or right after: when to do it
| Where you are | What to do |
|---|---|
| Still testing the idea, no real data in it | Keep prompting. Nothing to protect yet. |
| About to launch to real users | Get the review now. Closing holes before the data arrives is the cheapest this will ever be. |
| Launched in the last few weeks | Get the review this month. Every week adds records to an app nobody has checked. |
| Paying users, payments live, growing | Review, then plan the rebuild. You are running a production app on a prototype. |
Before launch is best for a plain reason: nothing has been exposed yet, so there is nothing to report, no keys to rotate in a hurry and no users to move. A hole closed before anyone's data is behind it costs you a fix. The same hole found after a leak costs you the fix plus the leak.
Right after launch is the next best. If you are already live, do not wait for the "right moment" between features. The trigger is not revenue or user count. It is the first real person's data.
What you can check yourself today
While you line up a review, the ten-minute versions:
- Log in as a normal user, open one of your records, change the ID in the address bar and reload. If you see someone else's data, stop taking signups until it is fixed.
- Press F12, use the app, and look through the network traffic for anything that looks like a secret key.
- Open your database dashboard. Does every table have access rules you understand?
- Export your data and try restoring it to a copy.
The full walkthrough is in Is Your Vibe-Coded App Safe With Real Customer Data?.
How Rootech fits in
We are a Melbourne software studio working Australia-wide, with delivery outside Melbourne done remotely. We rebuild apps made with AI builders and AI coding tools so they are ready for real users, and it starts with the review described above. On every build the client owns the code and the accounts stay in their name.
If you are about to launch, or just have, send us what you built at /#contact. You can also read what a cleanup involves end to end in Vibe Code Cleanup in Australia: What It Is and How Rootech Helps.
Frequently asked questions
Is a vibe-coded app really less secure than one a developer wrote?
Developers write holes too. The difference is review: professional code is usually read by a second person and covered by tests before it ships, and most vibe-coded apps have had neither. The research above found the same specific holes over and over, which is what unreviewed code looks like.
My app only has a few users. Is it worth doing yet?
If those users' real details are in it, yes. Attackers find open databases with automated scans, not by reading about you. A small app is not harder to find than a big one.
Does a security review mean rebuilding the whole app?
No. The review tells you what is needed. If the foundation is sound, a focused pass on permissions, keys, backups and monitoring may be all it takes.
How much does it cost and how long does it take?
We quote a fixed price after reviewing the prototype, before any work starts. Most Rootech builds ship in two to four weeks, and the prototype keeps running throughout.
Can I keep using Lovable, Bolt or Cursor afterwards?
Usually yes. An app with tests and a review step is a safer place to use AI tools, because the tests catch what a bad prompt would otherwise ship.
Is this the same as a penetration test?
No. A penetration test is a person actively trying to break in. Our review reads the code and configuration to find the holes directly. If a customer or insurer asks for a penetration test, do it after the review's findings are fixed, so you are not paying a tester to find what was already on the list.