Codex runs in your terminal, in your editor or in the cloud, where each task gets its own workspace and ends in a diff you can commit or turn into a pull request. A pull request you merge without reading is still code nobody has read.
I read it. You get a ranked list of what to fix, in plain English, with a fix prompt for each finding you can hand to Codex as its next task.
1Line 5: Logged in is not the same as allowed; the query never mentions the user, so any customer can read any invoice by changing the id.
Common problems in OpenAI Codex apps
Not every app has all of these. The report tells you which ones yours has, where, and what to do about each.
Critical
Logged in, therefore allowed
The handler checks that there is a session, then loads whatever record id the request names. Any customer can read any other customer's data by changing a number in the URL. I check every lookup for a where clause that includes the user.
High
Tests that test the mock
A task that ends with green tests looks finished. Some of those tests mock the function under test, or assert only that the mock was called, so they pass whatever the code does. I tell you which tests check something real.
High
Validation at the form, not at the API
The form has a zod schema. The route handler behind it reads the body and writes it to the database as it came. Anyone who skips the form skips the schema. I list each entry point and what a caller can change through it.
Medium
Every task brought its own helper
Each task was told to make one thing work and did, with its own formatDate, its own fetch wrapper and its own idea of how errors look. I map the duplicates and say which one to keep.
Medium
Errors that return 200
Handlers that catch an error and respond with status 200 and an error field in the body, so the frontend shows success and the monitoring sees nothing. I trace each one to what the user ends up believing.
Low
TODOs that became permanent
A task told to skip rate limiting for now left a comment saying so. Nobody came back. I collect every TODO and FIXME the agent left and rank them by what happens if they stay.
Treatment plan
How it works
1
Book a check-up
Choose a tier and tell me where to send the report. Add your app's address if it is live. Two minutes, no sales call.
2
Show me your code
Pay by card, then give me read-only access to your code on GitHub (you paste in a key I send you) and let me into your app with test accounts. Never done that? I send you step-by-step instructions for your tool. I measure your app and confirm it fits before I start.
3
Get your diagnosis
Ready to launch, or not yet. Problems ranked, critical ones proven on your live app, each with a prescription for your AI tool. A PDF by email, then 7 to 30 days of questions by email.
4
Fix it, then a free follow-up
Fix the problems yourself, with your AI, or let me do it with the Fix Pack. Tell me within 30 days and I check once more that they are really gone.
No idea how big your app is? That’s normal, and you never have to count anything. Book the review you want; I measure your app before I start and tell you. If it doesn’t fit, you choose: upgrade, narrow it down, or a full refund. “Lines of your code” means the code you and your AI wrote, not the packages and UI kit your tool installed; many apps built with Lovable, Bolt or v0 fit the Safety Scan.
Start small, upgrade later. The Safety Scan’s price counts in full toward a Launch Review or Fix Pack for the same app within 30 days of your report.
Why this price? Software does the first pass, so you pay for the part only a person can do: checking every finding in your code, proving the critical ones and deciding what matters for your app. I take at most 3 Safety Scans and 2 Launch Reviews (or one Fix Pack) a week; if this week is full, yours starts the next.
Prices in US dollars, paid once, no sales call and no surprise invoice: anything extra is your choice, told to you before I start. The clock starts when I can see your code. Reports arrive as a PDF by email, in English, or German if you prefer.
Need a few things fixed after a scan? I fix by the hour at $199, three hours minimum, with the list agreed before I start.
Bigger than 50k lines, or a real mess? Rescue jobs are quoted, from $3,000, no obligation. Email me.
Not for you if
you have a team of engineers who already review each other’s code,
your app’s backend isn’t JavaScript or TypeScript (Python Replit apps, for example), or it is over 100k lines,
you handle regulated data (health records, banking, children), or need a pentest, certificate or compliance sign-off. Taking card payments through Stripe is fine.
A review lowers your risk a lot. It can’t guarantee that your app is secure, and nobody honest will say otherwise.
OpenAI Codex questions
How do I share code written with Codex?
Codex works on a Git repository, locally or in a cloud environment connected to GitHub, so the code is usually on GitHub already. Give me read-only access (I send you a key to paste in). If it only lives on your machine, push it first.
Codex already reviews my pull requests. Why a human?
An automated review is built around the diff in front of it. It will rarely tell you that the route added today has no auth because auth was never set up for that part of the app. I read the system, then the diff, and prove the critical findings against your live app with test accounts.
Can I apply the fixes with Codex, and will you look at my AGENTS.md?
Yes to both. Each finding comes with a fix prompt that names the files and the change, written so an agent can act on it without guessing. Read the diff before you merge it. Codex reads AGENTS.md before it starts work, so the rules at the end of the report are written to drop straight in.
Why not just ask my AI to review it?
Do, and I will even give you the prompt. You will get a long list. What your AI cannot do is tell you which items are real, prove them against your live app, see your database and hosting settings, or check that its own fixes worked. If you cannot read code, that is the part you cannot do yourself either. It is what you pay me for.
I'm not a developer. Is this for me?
Yes, that is exactly who it is for. If you built an app with AI and cannot tell whether the code is any good, I check it and explain what matters in plain English. You never need to understand the code to act on the report: every problem comes with a prompt you paste into your AI tool. And you do not need to know what to ask. The report answers the questions you would have asked if you knew how.
How do I show you my code?
Most AI builders can push your app to GitHub, a free site that stores code. You give me read-only access by pasting a key I send you into your repository settings, which means I can look but not change anything. Never done it? After you book, I send you step-by-step instructions for your tool. A zip file works too. When the report is delivered, remove my access again; I remind you.
Is my code safe with you?
Every booking comes with an NDA, automatically. I never sell, publish or pass on your code. The only other party that sees it is the AI provider that helps with the first pass, under terms that do not allow it to train on your code. I never need your passwords or your live database, and I delete my copy within 7 days of delivering the final report.
From $149 for the first 10 customers (10 left). Ready within 2 business days. If the Safety Scan finds no problem rated medium or worse, you pay nothing.