Security scanners for AI-built apps read configuration and code, so they catch row level security being off, secrets in the bundle and known package vulnerabilities. They cannot catch business logic flaws, data leaking through legitimate features, or sequence and timing problems, because each depends on what your app is meant to do.
1. What the free scanners genuinely do well
2. The three classes no scanner finds
3. Class one: business logic
4. Class two: leaks through legitimate features
5. Class three: state and timing
6. The two-account test, extended
7. Proving a fix actually worked
8. When you need a person rather than a tool
9. How the Vibe Check works
10. Frequently asked questions
There is now a good supply of free checklists for AI-built applications, several of them written by developers with real SQL in them, and they are worth running. This article assumes you already have. It covers the part that comes after, which nobody writes about, because scanning is easy to automate and the rest is not.
About Wow Labz. Wow Labz is an AI-native custom software development company based in Bengaluru, India. Since 2011 it has shipped 400+ products across 15+ years, won 30+ awards, and touched 100M+ lives. It holds a 5.0 rating across 23 verified Clutch reviews and is ISO 27001 certified.
What the free scanners genuinely do well
Start with the honest version: run them. They take minutes, they cost nothing, and they close the holes that account for most published incidents in AI-built software.
Anything that is a fact about your configuration is machine-checkable, and a tool will check it faster and more reliably than you will. That includes row level security switched off on a table, a policy written as true for everyone, a service key sitting in the browser bundle, missing security headers, known package vulnerabilities, exposed schemas, a committed environment file, and privileged helper functions that bypass your policies.
These matter because they are common. A Q2 2026 scan of 23,764 deployed applications, published by securityscanner.dev, found 7% of apps built on popular AI tools had databases anyone could read, against 0% for a Y Combinator control group. The Cloud Security Alliance reports Veracode finding that 45% of AI-generated code samples introduce OWASP Top 10 vulnerabilities, with no improvement across testing cycles.
So the scanners earn their place. The problem is what happens next, which is that a clean scan feels like an all-clear and is not one.
The three classes no scanner finds
A scanner reads configuration and code. It cannot read intent. Every flaw below depends on knowing what your application is supposed to do, which means no tool can find it without being told, and nobody has told it.
The same split in table form:
| Scanner catches | Scanner misses | |
|---|---|---|
| What it reads | Configuration and code as written | Intent, sequence and business meaning |
| Typical findings | RLS off, permissive policies, secrets in bundle, missing headers, vulnerable packages, exposed schemas | Business logic flaws, leaks through legitimate features, state and timing problems |
| How it is found | Automated, minutes, free | By a person using the app adversarially |
| Who it hurts | Everyone equally, and it is well documented | You specifically, in ways that depend on your product |
The distinction in one line. A scanner can tell you a door is unlocked. It cannot tell you that the door should not exist, or that the person walking through it is allowed in the building but not in that room.
Class one: business logic
Every request is authenticated. Every request is authorised. The user is allowed to do each individual thing they did. The problem is the sequence, or the arithmetic, and no policy engine has an opinion about either.
-
Upgrading without paying. A plan upgrade where the entitlement is granted before the payment is confirmed, so cancelling at the right moment leaves the account upgraded and unpaid.
-
Refunding twice. A refund endpoint that does not record that a refund already happened, so calling it twice refunds twice. The caller is entitled to request a refund, which is exactly why nothing blocks it.
-
Reusing a one-time coupon. A discount code checked for validity but not for prior use by that account. Each individual redemption is legitimate.
-
Spending credits you do not have. Credits deducted after the work is delivered rather than before, with no lock, so parallel requests all pass the balance check.
This is the class that costs money directly, and it is close to invisible in logs because nothing errors. The requests all return 200.
Class two: leaks through legitimate features
The data layer is correctly locked down. Row level security is on, the policies are scoped, and a scanner gives the app a clean bill of health. Then a feature hands the data over anyway, because the feature runs with permissions the user does not have.
-
Search across tenants. A search index built without tenant scoping, so querying returns titles, snippets or counts belonging to other accounts even though opening the record is correctly blocked.
-
Exports that ignore the filter. A CSV or PDF export generated server-side, frequently with an elevated key for convenience, which ignores the filter the interface applied.
-
Error messages that confirm existence. A not-found message that differs from a not-allowed message, letting someone confirm which records exist. Minor alone, useful in combination.
-
Counts and totals. Aggregates computed across the whole table rather than the caller’s rows, so a dashboard quietly reveals the size of the business next door.
The pattern is consistent: a privileged server-side process doing something helpful, built by a model that was asked to make the feature work and not asked who should see the output.
Class three: state and timing
Each step is valid. The order is not. These are found by doing things in sequences nobody anticipated, which is precisely what a scanner cannot do because it does not know what the intended sequence was.
-
Skipping a verification step. Navigating straight to the step after email or identity confirmation, because the later step checks that you are logged in rather than that you are verified.
-
Re-using a one-time link. A password reset or invitation link that is not invalidated on use, or not on expiry, or not when a second one is issued.
-
Acting after access is revoked. Sessions, tokens or cached permissions that survive a removal from a team, so a departed user keeps working for as long as the token lives.
-
Racing a check against a write. Two requests arriving together that both pass a check before either writes. Classic in anything involving a balance, a quota or a seat count.
The two-account test, extended
One manual procedure finds a large share of all three classes. It takes about twenty minutes, needs no tooling and no security background, and most people stop after the first quarter of it.
Create two real accounts. Account A holds data worth stealing. Account B is the attacker, and should have a paid plan if you sell one.
-
Swap the identifiers. Logged in as B, paste a URL or record identifier belonging to A. Then try the API directly, outside the interface. Pass means nothing comes back. Fail means anything at all comes back.
-
Use every feature. As B, exercise every feature: search, filter, export, share, print, the API. Look for any trace of A’s data in the output, including counts, titles and snippets. This is the step people skip and it is where class two lives.
-
Test privilege. Reach for things B should not have: admin URLs, other people’s settings, billing pages, features belonging to a plan B has not bought. Guessing a URL counts as an attack and should be treated as one.
-
Do things in the wrong order. Skip a step, go back, resubmit, reuse a link, cancel then act, apply the coupon twice, open two tabs and submit both. This is where class three and most of class one surface.
The rule that makes the test work. Never test by looking at the screen. Test by looking at the response. An interface that hides a record is not the same as a server that refuses to send it, and the difference is the whole vulnerability. Open the network tab and read what actually came back.
Proving a fix actually worked
The gap that turns a one-time audit into a recurring problem. Most checklists tell you how to fix something. Very few tell you how to know it is fixed, and an unverified fix is a belief rather than a change.
-
Re-test the exact thing that failed. Confirming through the interface proves the interface changed. Re-run the original request directly against the API with the same credentials that worked before.
-
Re-test the neighbours. A policy added to one table rarely covers the view, the function, the export path or the sibling table created in the same session. Fixes tend to be narrower than the flaw.
-
Write the test down. Two tests per table, written once: an anonymous request returns nothing, and a second user’s request returns nothing. If you write nothing else, write these.
-
Re-run them before every release. Most regressions arrive with the next feature, generated by a model that has no memory of the policy you added last week. Without a test, you will not find out.
When you need a person rather than a tool
Four situations where self-testing is not enough, stated plainly rather than as a pitch.
-
You hold money, health or identity data. If a wrong answer costs someone money, access or privacy, the cost of missing something exceeds the cost of having it checked properly.
-
Someone is about to ask for an audit. Enterprise buyers, insurers and regulated partners increasingly ask for evidence. Your own testing is not evidence, however thorough it was.
-
Nobody can explain how it works. Debugging an app nobody can explain is a different exercise from reviewing one you designed, and it is the normal situation after a few months of generated features.
-
You found something and cannot tell whether it was used. You are then handling an incident rather than a vulnerability, and the clock on disclosure obligations may already be running.
How the Vibe Check works
We run the Vibe Check because of exactly the gap this article describes. Automated scanning is part of it and it is the smallest part. Senior engineers spend most of the 72 hours doing what a scanner cannot: using the application adversarially, in the order nobody intended, looking for the three classes above.
What comes back is a signed report with findings ordered by reachability rather than by severity label, each with the request that demonstrates it, plus a fixed-price plan for closing them. It is free, it starts with an NDA, and access is read-only.
If you have already had findings and are deciding what to do with them, our guide to triaging a failed security audit covers the ordering and the patch-or-rebuild decision. Where remediation turns into rebuilding, custom software development covers how we scope that.
Clean scan, still uneasy?
That instinct is usually right, because the scan only covered the part a machine can read. Senior engineers audit your app in 72 hours, test the three classes above by hand, and hand back a signed report with each finding demonstrated rather than asserted. NDA first, read-only access, twenty founders a week. Get a free Vibe Check and find out what the checklist missed.
Frequently asked questions
What do security scanners miss in AI-built apps?
Three classes: business logic flaws such as upgrading a plan without paying or refunding twice, data leaking through legitimate features such as search or exports that ignore tenant scoping, and state and timing problems such as skipping a verification step or reusing a one-time link. Each depends on what the application is meant to do, which a scanner cannot know.
Is a clean security scan enough before launching?
No. A clean scan means the configuration is correct, which is necessary and not sufficient. It says nothing about whether a paying user can reach another tenant’s data through your search feature, or whether your refund endpoint can be called twice. Run the scanners, then test the logic by hand.
How do I test my app for business logic flaws?
Use the two-account test. Create two real accounts, then as the second account try swapping identifiers, exercising every feature while looking for the first account’s data, reaching for admin and paid features, and doing things in the wrong order. Read the network responses rather than the screen, because a hidden record is not a refused one.
What is the difference between authentication, authorisation and business logic flaws?
Authentication asks whether you are who you say. Authorisation asks whether you may perform this action on this object. Business logic asks whether the sequence of permitted actions produces a result the business intended. Scanners handle the first two reasonably well and cannot reason about the third.
How do I know a security fix actually worked?
Re-run the exact request that failed, directly against the API rather than through the interface, using the same credentials that previously succeeded. Then test the neighbouring tables, views and export paths, because fixes are usually narrower than the flaw. Write the test down and re-run it before every release.
When should I pay for a professional security audit?
When the app holds money, health or identity data, when a customer, insurer or regulator is about to ask for evidence, when nobody on the team can explain how the application works, or when you have found something and cannot establish whether it was already exploited. The last case is an incident rather than a vulnerability.