AI coding tools write a lot of code quickly, and most of it works. The problems tend to sit in the places a demo never touches: who can read what, where the keys live, and what happens when something fails. These nine come straight from the patterns that recur in readiness reports. Each takes minutes to check.
1. Keys in the code
Search the repository for anything that looks like an API key, including in old commits. A key that was committed and then deleted is still in git history, so deleting the line does not help. Rotate the key first, then move it to an environment variable.
2. The database admin key in the browser
If you use Supabase, look for SUPABASE_SERVICE_ROLE_KEY anywhere a browser can reach. That key ignores every row-level rule you wrote. In client code, it makes your whole database readable and writable by anyone who opens the developer tools.
3. An .env file that git can see
Check that .env is listed in .gitignore. Commit a .env.example with the variable names and no values, so the next person knows what to set.
4. Records anyone can read by changing a number
Sign up twice, as two different people. As the second person, open a record that belongs to the first by changing the identifier in the address or the API call. If it opens, you have the most common serious flaw on the web. We wrote about it in The bug on line 7.
5. Queries built by joining strings
Look for SQL built with + or template strings around a value from a form or the address bar. Use parameterised queries instead. This is how a search box becomes "read every row in the table".
6. Text from users written into the page as HTML
Search for innerHTML and dangerouslySetInnerHTML. If what goes in came from a user, it can carry a script that runs in someone else's session. Use text instead of HTML, or sanitise it first.
7. An API that accepts calls from any website
A CORS setting of origin: '*' together with credentials lets any website make signed-in calls to your API on your visitors' behalf. List the origins you actually use.
8. No tests on the path that earns money
You do not need a coverage target. You need one real test over sign-up, checkout, or whatever your business runs on, so the next AI edit cannot break it silently. AI edits are large, and one test is the difference between noticing on Tuesday and noticing from a customer.
9. Failures nobody can see
Search for empty catch blocks and network calls with no timeout. The first hides errors you will spend days looking for. The second lets one slow service hold every request open until the app stops answering.
Or let us check
An AIpento readiness report runs these checks and many more against your repository, ranks what it finds by severity, and shows the file, the line and why it matters. It never runs your code. See a complete sample report, or see the plans.