API Maker

The backend engine for the AI era

Security reportEpisode 27 12:24 20 chapters November 2026

APIs Security Report: Find Data Leaks and Fix Them in One Click

API Maker reads your whole account and tells you where a person could reach data that is not theirs, with the proof, then fixes it one click at a time. Scan Calm Home after a busy sprint: a score out of 100, how the scanner thinks, fixes for groups, system APIs, person tokens and row scoping hooks, a PDF for the review, automatic scans, and the report in Git for your AI assistants.

In the playlists The complete course Security & access

Chapters

What you will learn

  • Scan the account: a score out of 100
  • Two gates, one report: how the scanner thinks
  • A finding in detail, with its proof
  • Fixes in one click: narrow a group, require a person token, a row scoping hook
  • Skip a finding with a reason
  • The score scan after scan, export to PDF, automatic scans
  • The report in Git, for AI assistants
In the docsAPIs Security Report & Actions
Transcript1803 words · click a time to play from there

Intro

API Maker reads your whole account. It tells you where a person could reach data that is not theirs. Then it fixes it, one click at a time.

Security reviews come too late

Security reviews usually come too late. They are manual, late, and rare: a spreadsheet before a release, when somebody has the time. Meanwhile, one application token reads every row of a table. A wildcard grant reaches the tables you add next year. And a hook that looks right filters only one API out of four. Backends written fast, by people or by AI assistants, ship with these holes. API Maker scans the account on every change.

It shows the proof. It applies the fix through the same checked routes. And it keeps every decision in Git. In this video: a weak account, the scan and its score, how the scanner thinks, one finding in detail, the fixes, with the score going up, export and print, and automatic scans, kept in Git for your AI assistants.

The weak account

Here is Calm Home after a busy sprint. The customer app got a login table in PostgreSQL, with the auth provider app_login, and its own group. A back office tool got Execute plain query and Get secret keys. Countries became public. The hook of the support tickets was switched off while debugging. And the Default group, as in every new account, still reaches everything. Nothing here looks dangerous on its own.

But the token of the customer app reads all 180 orders of 59 customers.

Scan and score

In the Security panel, open APIs Security Report and Actions. Before the first scan, the page lists the six questions it asks: row gate, public access, code, API gate, auth providers, and hygiene. For this demo, we keep the scan to what API Maker already knows. Then we scan the account. The score is zero out of a hundred, grade F: unprotected. There are open findings, several of them critical, across six instances and twenty four collections.

Every open finding takes points off a hundred: twenty for a critical, ten for a high, four for a medium, and one for a low. Below, you see where the risks are: one line per category, and the clear ones are ticked. A severity opens its findings, with the summary folded to one line. One click clears the filter. And the findings are grouped by category, the worst first.

How it thinks

Every request carries two identities, and the report checks both. The API gate is the API user token, in x-am-authorization. It answers: which APIs can this application call? The row gate is the token of the person, in x-am-user-authorization. It answers: of the rows this API reaches, which ones belong to this person? The row gate is the heart of the report. Without a filter, a caller allowed on a table reaches every row of it.

Here is a simple example. Diya and Kabir both shop with the customer app. The phone of Diya asks for orders with two tokens: the token of the app, and her own. The API gate lets the app read orders, but nothing limits the rows. So Diya changes one number in the request, to customer five, and she reads the order of Kabir, with its number and its total. With a row scoping hook, the same request answers nothing.

The job of the report is to find these tables before Diya does. A scan is a pure function of a snapshot. It reads the groups, API users, auth providers, settings, secrets, schemas, and pre-hooks. It works out which group reaches which API of which collection. It guesses the column that names the owner of a row. That can be a relation to the users table, a column named after it, or a common name like user id.

It reads every pre-hook, in any of the eight languages, without running it. It checks four families of APIs, and checks that each filter is built from the person. Then dozens of rules write the findings, and more rules read the code itself. All of it is Rust, and the scan takes well under a second.

A finding

Now open a finding, here orders, in MySQL. The top shows the severity, the category, the target, and a link to its page. Why it matters is said in plain words. Then come the APIs it can reach, grouped by what they read, change, or delete. And the evidence: the groups that allow them, a person token that is not required, and the owner column, customer_id. It also says how sure it is: the column is named after the users table of app_login.

At the end, How to fix it sits above the actions that apply the fix.

Fix it

Now, the fixes. Every action goes through the same routes as the admin panel. That includes checks, the TypeScript compile, and Git sync. And the account is scanned again right after.

Fix: narrow a group

Here is the first critical finding. Default reaches every instance, database, collection, and API, even the tables added next month. The dialog offers two ways. Freeze writes what the group reaches today as exact grants, then turns the flag off. Turn the flag off only keeps the exact grants, and Default has none. The customer app has its own group, so Default does not need the data. Apply, and scan again.

Several findings close at once, and the three flags below it go the same way.

Fix: system APIs

Default also reaches every system API, even Execute plain query and Get secret keys. This flag cannot be frozen, so it goes off, and the score goes up. The back office group names the two risky ones by name. Revoke these system APIs lists them as checkboxes, and the group keeps every other grant. Apply, and the score goes up again.

Fix: skip with a reason

Countries is public on purpose. The app shows it before anybody signs in. Skip asks why that is fine. The reason is kept with the report, and committed to Git. The finding leaves the score, and the score goes up. The Skipped tab keeps it, with the reason, who skipped it, and when. Open it again, and it comes back, with its twenty points. We skip it once more, and move on.

Fix: person token

Now the row gate. Orders belong to people, but their APIs need no person token. Require a person token writes the collection settings of orders, with app_login. Apply it. Now every API of orders needs a person, next to the API user. The scan tells the next truth: a person is known, but nothing limits the rows to theirs. That is critical, so the score drops. The report counts risk, not effort.

Fix: row scoping hook

The evidence shows a gap. The token of app_login carries the email and the id, but not customer_id. One click adds customer_id to the select of the provider. Then, add a row scoping pre-hook. On its Settings tab, the owner column and the identity in the token are picked for you, and the safe behaviour is on. The Code tab holds the generated hook. You can edit it before you save it.

Apply, and the score climbs. The generated hook follows one standard. Inside calls from custom APIs and schedulers pass through. No person means no rows. The owner is merged with dollar and, never replaced, so a caller cannot widen the rows. Every write is stamped with the owner. And aggregate and distinct are refused. It fails closed. Diya signs in to the customer app. With the token of the app and her own, orders now answers with her two orders, and nothing else.

Fix: activate a hook

Support tickets has a hook, Only my tickets, that is switched off. The report has already read its code. It reads the person and filters get all, but not query. It stamps nothing, refuses nothing, and merges with an object spread. Activate it, and the score goes down. The hook is graded on what it does, not on its name or its comments. That is how a hook that looks right gets caught, whether a person or an AI assistant wrote it.

The finding lists the APIs that the hook leaves open. Replace with a hook covering every API installs the generated hook, and the score climbs again.

Links to pages

Some fixes are yours to make, like a groups column for app_login. The link opens the page of the target in a new tab, so the report stays where it is.

The score

The trend of the score tells the story. It was zero at the first scan. It went up with every flag, grant, and token. It went down once, when the old hook of support tickets was switched on and graded. It went up again with the generated hooks, and reached grade A once the last flags of Default were off.

Export and print

The report can also leave the page. The download menu prints it, with or without the skipped findings. It also downloads report.md, report.yaml, or report.json. Click Print, then Save as PDF in the browser dialog. Page one is the summary: the score, the severities, what to fix first, and what was scanned. The findings follow, with their evidence and their fix. The Markdown file has the same content, for a ticket or a pull request.

Automatic scans

While a local client is connected, the report stays fresh by itself. When a scan is due, one count and one date per model show if anything changed. Only a change starts a scan. Only a different report is saved, and it reaches your disk through the local client. Committing it is still your choice, on the Git page. The next scan is due a hundred times the length of the last one later, between one and thirty minutes.

Git and AI

In the repository, the report is a folder of four files. report.md, report.yaml, actions.yaml with every skip and its reason, and a README for people and AI assistants. It says: read only. Work on the findings, not on the report. Respect the skips. An assistant reads report.md, and fixes the hook in its own file. Then the next scan grades its work. API Maker reads only one thing back: the skipped findings.

Best practices

Here are the best practices. Keep the local client connected, so every change is scanned. Use no broad flag in a production group. Ask for a person token on every table that belongs to people, with the generated hook. Skip a finding only with a reason that holds. And keep the report in Git, next to the code.

Recap

That was the APIs Security Report. One scan gives one score. Every finding has its proof. Fixes take one click, through the routes of the panel. There is a PDF for the review, and the report in Git for your AI assistants. Next, the other security features: encryption, secrets, origins, and two step login. API Maker. Build faster, spend smarter, deliver better.