I have run read-only security audits on 7 live applications. All of them were small apps built by one or two people, with real users and no security budget. The same five problems kept coming up, and none of them are unusual.
I do not name any application or give reproduction steps. Some of what I found is still being fixed, and a writeup that points at live holes would help attackers. The vendors whose products are involved document each of these categories publicly.
Row-level security is on and enforcing nothing
This was the most common one in my audits, and it is easy to miss because the dashboard shows green.
Enabling row-level security on a table means no rows are returned unless a policy permits them, as the vendor documentation says. Many people misread that sentence. They write a policy to make the app work again, and the policy they write is usually too broad. A policy written without a role clause applies to every role, including the anonymous one, and the anonymous key ships inside the page bundle by design. Anyone can read it out of view-source.
Your application may filter rows correctly, but an attacker skips your application and calls the database API directly with a key you published.
For every policy, check which roles it applies to. If you did not specify a role, it applies to all of them.
Paid API routes with no authentication
Small apps often have server routes that call a metered service, such as a language model or a maps API. Developers skip authentication on these routes because during development they are the only ones who know the URL. That URL is in your JavaScript bundle, where anyone can read it.
An unauthenticated route that calls a paid API lets anyone run up your bill. If someone calls it in a loop, your card pays for their traffic, and without alerts you find out from the invoice.
Do the console fixes first. Set spend caps and budget alerts on every provider, and add referrer or IP restrictions to any key that supports them. They cost nothing and limit the damage right away, while the code fix waits for a session where you have time to do it properly.
An auth cookie anyone can set
This one starts when a developer writes a placeholder login check during the build and never replaces it. The app sets a cookie on login, and the check asks whether the cookie exists, or whether it equals a known constant.
Anyone can set a cookie. If your check is a string comparison against something a client controls, the gate does nothing, and it is usually protecting whatever the developer considered too sensitive for the public routes.
To test it, open a private window, set the cookie in the console, and load the protected page. If the page loads, anyone can get in the same way.
Sequential identifiers as authorisation
Say a record is at /thing/41 and a logged-in user requests /thing/42. Many small apps never check whether that user may see it.
In an app where the identifier is also the only thing separating one user's data from another's, incrementing an integer is the entire attack. I found this in voting, ordering and messaging features, where the developer thought about the interface and never about the endpoint.
Check authorisation on the server for every request, whether or not the user knows the identifier.
Input that becomes markup
This is text a user typed, such as a display name or a review, rendered back out as HTML. Injection is on the OWASP Top Ten, and I still found it.
Frameworks escape by default now, which has made this rarer. Where it survives, it is because someone deliberately opted out to get formatting working. Search your codebase for whatever your framework calls "render this as raw HTML" and check where each one gets its input.
How do I run one of these audits?
I keep every audit read-only. The output is a list, and the fixes happen later in a planned session.
Every finding gets three parts, and I do not count it as a finding until it has all three.
SEVERITY what an attacker gets, in one sentence
REPRO the exact request or steps that demonstrate it
FIX the specific change, not "add validation"Keep two lists, one for findings you verified against source and one for findings you suspect but could not confirm. If you mix them and one confident claim turns out to be wrong, the reader discounts the four that were right.
Why does the developer miss these?
Each of these problems sits in code that works. The developer tested that the features work, and that testing does not show what an attacker could do.
A second reviewer helps because they did not write the code, so they check what it does instead of assuming they know.
If you run a small app with users and want someone else to look at it, I will do it for free: elijah@purcellventures.co.
Part of the guide AI integration for small businesses.