Security testing and penetration testing

Find the holes before someone else does.

I test your app, API or AI feature the way an attacker would. You learn what is open, how serious it is and what to fix first. If you want, I fix it too.

Test types
Black, grey and white box
Surfaces
Web, API, mobile, cloud, AI
Standard
OWASP WSTG and ASVS, CVSS 4.0
Environment
Isolated copy with synthetic data
AI
Anthropic's Cyber Verification Program
Delivery
Report, walkthrough and retest pack

01 · Three ways to start

How much should I know before I begin?

The answer decides which attacker the test imitates, and what kind of flaws it finds.

  • Black box

    From the outside

    I get only the address, like a stranger on the internet.

    • Shows what the world can see and take today
    • All you provide is written permission

    Before launch, or as a first check.

  • Grey boxMost common

    As a user

    I get accounts for each role, like a customer with bad intentions.

    • Finds flaws in who can see and do what
    • You create two test users per role

    Apps with logins, roles or payments.

  • White box

    From the inside

    I get the code, the database and the cloud setup, and test all three levels.

    • Finds what cannot be seen from outside
    • You give read access

    Personal data, money, or before an investor looks at you.

02 · Fig. 1 · A finding, as it appears in the report

F-04 · ExampleHigh · CVSS 4.0: 7.1

A logged-in user can read other customers' invoices

Where
GET /api/invoices/{id}
Cause
The server checks that you are logged in, but not that the invoice is yours.
Fix
An ownership check in the query, and a test that locks it in.
Confirmed
Against the running system
Retest
Closed

Why a developer should test your app

I have built apps, backends and AI systems for more than ten years. I know where the shortcuts get taken, because I have watched them being taken.

That is why the report explains the cause and the fix, not just the symptom. And it is why I can fix the flaws myself, if you have nobody to do it.

03 · How it works

Three passes over the same system.

When I get access to the code, I do not test in production. I set up an isolated copy of the system with made-up data, and in there everything is allowed. Then I go through it three times, from three different starting points.

  1. 01

    Scope and agreement

    Which systems, what is out of scope, and written permission from whoever owns the system.

  2. 02

    The test environment

    A copy of the system that runs only on my side, with made-up users for every role. No real personal data and no contact with production.

  3. 03

    Pass 1: from the inside

    The code, the database and the access rules are read, and automated scanners cover the known issues. Earlier findings are retested.

  4. 04

    Pass 2: as a user

    Every function and every API call is tried with every role: logged out, normal user, read-only, admin at another customer, and someone whose access has been removed.

  5. 05

    Pass 3: blind

    An independent pass with no code and no knowledge of the findings, only test accounts and the address. It is run by AI models for penetration testing, shows what an attacker finds unaided, and catches what the first two missed.

  6. 06

    Confirmation

    Every finding must be shown against the running system. What cannot be reproduced is struck.

  7. 07

    Report and walkthrough

    You get the report and a meeting where you can ask about everything.

  8. 08

    Retest

    All the checks sit in a pack that can be run again with one command. When the flaws are fixed, it shows it.

04 · AI in the testing

Dedicated models for penetration testing.

I use AI in the testing. I am approved in Anthropic's Cyber Verification Program, which gives access to dedicated Claude models for penetration testing in controlled environments. They search wider and faster than one person can, and test the same functions with every role without getting bored.

The models never get outside the test environment. They work against the isolated copy with made-up data, not against production. Every finding they report, I confirm myself against the running system before it enters the report.

05 · The delivery

A report both management and the developer can use.

  • Executive summary

    One page: overall risk, what to fix first and why, and what held.

  • Every finding can be verified

    Step-by-step reproduction, cause, impact and recommended fix. Each finding also says how it was confirmed: against the running system, against the database or in the code.

  • Classified by the standards

    OWASP category, CWE and CVSS 4.0 on every finding, with an explanation of what the score assumes.

  • What held

    The controls that worked are in the report too. That tells you what not to touch.

  • Retest pack

    All the checks in one place, ready to be run again by me or by your team.

  • Limitations

    What was not tested, and what should be checked against production before a finding is closed.

06 · In depth · The full finding

Every finding can be run again.

F-04 · ExampleHigh · CVSS 4.0: 7.1

A logged-in user can read other customers' invoices

Status
New
Confirmed
Against the running system, over HTTP
Where
GET /api/invoices/{id}
Category
OWASP API1:2023 (BOLA) · CWE-639
Vector
CVSS:4.0/AV:N/AC:L/AT:N/PR:L/UI:N/VC:H/VI:N/VA:N/SC:N/SI:N/SA:N
Reproduce
Log in as customer A. Fetch your own invoice. Change the id to an invoice that belongs to customer B. The server returns the whole invoice.
Cause
The endpoint checks that the user is logged in, but not that the invoice belongs to the user.
Impact
Every invoice can be fetched by counting through the ids: names, addresses and amounts for all customers.
Fix
An ownership check in the query on the server, and the same check on the other endpoints that take an id.
Retest
Check P-04 in the retest pack. It fails today, and passes once the check is in place.

The finding is a constructed example that shows the format. Real findings are never shared.

07 · In depth · Rules

Rules of engagement.

  • Written permission first

    No testing without written permission from whoever owns the system. Without it, testing is illegal, whatever the intent.

  • Third parties on their terms

    Vendors and cloud services are tested only within the rules they have set themselves.

  • Production is not touched

    The tests run in an isolated copy. In there everything is allowed, also what breaks things. In production I test only what we have agreed explicitly.

  • Made-up data only

    Every user, customer and address in the test environment is created for the test. No real personal data is used.

  • Staff are not tricked

    Social engineering happens only when it has been agreed explicitly.

  • The findings are yours

    I never mention them without consent, not even anonymously.

08 · Apps built with AI

Built with Lovable, Cursor or Bolt?

Apps built with AI tools have the same flaws as other apps, but some turn up more often. It is nothing to be embarrassed about. The tools build what you ask for, and security is what nobody thinks to ask for.

  • Anyone can read the database

    The rules for who gets to see what are not switched on. Anyone with the app can then fetch every user, every message and everything else stored.

  • Secret keys sit inside the app

    Keys for payments, email or AI services are shipped to the browser. Others can use them, on your bill and in your name.

  • Change a number, see someone else's data

    The app checks that you are logged in, but not that the data you ask for is yours.

  • The admin page is only hidden

    The button is gone for normal users, but the function behind it works for anyone who finds it.

  • AI with no cap

    An AI feature with no limits can be used by anyone, as much as they like. That can mean a large bill overnight.

  • No limits on logins and uploads

    Without limits, someone can try thousands of passwords, or upload files the app was never meant to accept.

In February 2026, an app built with Lovable was left open with data on around 18,000 users, because the database access rules were missing. In September 2026, TechCrunch reported Supabase customers exposing large amounts of personal data.

The Register, February 2026TechCrunch, September 2026

Common questions about security testing.

Which test should we choose?

If you are unsure, start as a user, with grey box. It finds the flaws that most often lead to leaks. A full test goes through all three: from the inside, as a user and blind.

What does a security test cost?

It depends on the type of test, the number of roles and how much the system connects to. You get a concrete proposal with a scope before anything starts.

Do you test in production?

Preferably not. I set up an isolated copy of the system with made-up data, and test there. That lets me try everything, also what breaks things, with no risk to your users. Findings that depend on the production setup are checked against production by agreement.

Do you use AI in the testing?

Yes. I am approved in Anthropic's Cyber Verification Program and use dedicated Claude models for penetration testing in controlled environments. They work only against an isolated copy with made-up data, and I confirm every finding myself before it enters the report.

Do you fix what you find?

Yes, if you want. The fixes can be done by me, by your team with my support, or by you alone. The retest happens either way.

What is the difference between a penetration test and an automated scanner?

A scanner finds known flaws in known software. It does not understand that one customer should not be able to see another customer's invoice. Those are the flaws behind the big leaks, and they are found by hand.

Is an app built with Lovable safe?

It can be, but often it is not out of the box. Lovable and similar tools build what you ask for, but the rules for who can read what in the database often have to be set up correctly on top. A test shows whether that has been done.

How often should we test?

Before launch, after larger changes to login, access or payments, and otherwise yearly for systems with personal data.

What if the app has already leaked?

Close the hole first, and find out what may have been taken. If personal data is involved, the breach must as a rule be reported to the data protection authority within 72 hours. Get in touch, and I will help with both.

Tell me what should be tested.

Write what the system does and what it is built with. You get a reply within a day, with a suggested test type and scope.