AI App Security Checklist: Test Who Can Read Candidate Data

Before your AI-built hiring app takes its first real CV, test who can read candidate data with an access acceptance sheet, allowed controls and saved evidence.

Ernest Bursa

Ernest Bursa

Founder · · 7 min read
Three startup colleagues review access decisions at a table, with an older man in a startupkit shirt and a laptop showing its plain back.

Your candidate form works. A test application arrives, the CV opens, and the hiring manager can leave a note. You could put the link in your next job advert and collect the first real résumé today.

One decision is still missing: who else can read that application? The demo shows that a permitted person can use the app. It says little about a candidate reading an internal assessment, a reviewer opening another job’s applicants, or a removed colleague keeping access to files.

Before real intake, write down who may read each part of an application. Create a synthetic record, check its data-delivery routes under allowed and forbidden identities, and save the results against the build you intend to launch. That gives your AI app security checklist a concrete permission test.

A working login leaves one question unanswered

The concern is concrete. In its analysis of Lovable-built sites, Crissy Field reports finding sensitive records accessible without login, including identity-document details. Its sample came from Common Crawl and was not representative. An email address in a returned record could be public or synthetic; the researchers did not establish misuse or show that AI caused the failures.

The study cannot predict what is wrong with your app. You can take its practical question seriously: what does a request receive before, or beyond, the access you intended?

Your builder may already offer tools to investigate that question. Lovable’s security documentation describes publish-time Quick scans and deeper code analysis, including authorization concerns. It also says those tools cannot guarantee complete security or replace thorough review. Keep the scans, then supply the business decision they need to support: this reviewer may read these fields for this job.

Authentication establishes identity. Authorization decides whether that identity may read a particular record. OWASP’s object-level authorization guidance calls for that check at each endpoint receiving an object ID. A hard-to-guess application link does not establish permission to its contents.

Write the access decision before testing it

Imagine you are hiring an engineer for your first product team. The application contains contact details, a CV and an internal assessment. The candidate should see the information your candidate portal promises them. An assigned reviewer needs the fields required for review. Someone helping with a different vacancy may have no reason to see any of it.

Ask the hiring owner to approve those distinctions before an engineer tests them. Otherwise, an unexpected result can become an argument over what the app was supposed to do. Be explicit about administrators, broader recruiting roles and any deliberate exceptions. The example below assumes a restricted vacancy; adapt it to your own rules.

Create one application in staging with a distinctive fake name, a dummy CV and a clearly marked internal assessment. Use records and accounts you control. Give the CV its own fake marker so you can check file access separately from access to the application page.

Copy this acceptance sheet and replace its expectations with approved ones:

Identity or state Expected access to the synthetic application What to inspect
Assigned reviewer Approved candidate fields; assessment only if intended Returned markers and field names. This is the allowed positive control.
Signed-out visitor No private application, CV or assessment The actual data and file responses before login.
Candidate account Its own candidate-facing fields; no internal assessment Response fields, including data hidden by the page.
Reviewer on another job No access unless a written rule grants it A direct request for this known application.
User in another company No private record or attachment from the first company The same resource request under the other company’s identity.
Removed team member No new reads once the defined revocation takes effect A fresh request after removal, with session state recorded.

A role name alone is too vague. “Reviewer can view application” leaves the assessment unresolved. Write the permitted fields beside the role, including any internal fields that must be absent. For a candidate account, specify whose application it can read as well as what it can see.

Follow the CV beyond the visible page

Start with the assigned reviewer. Confirm that the request returns the synthetic marker and approved fields. This positive control proves you are testing a working route and the intended record. If every identity receives an error because the endpoint is broken, the denied rows have proved nothing.

Then request that same synthetic record under each remaining identity. Have an engineer help inspect the network response where needed. A page can hide an assessment while still receiving it in JSON, so looking at the visible screen is insufficient. Keep real candidate information out of this exercise.

Follow the delivery paths your app actually has: application lists, detail pages, exports, CV downloads, alternate response formats and server functions. Include each route that returns the same private information through a different interface.

Kit provides a concrete example of why this matters. Its hiring application controller authorizes the detail request before serving either HTML or Markdown. Its application policy checks the account and access to the job posting. Kit’s team access documentation explains the broader role model; those rules still need specific checks wherever application data is delivered.

If your app uses Supabase, ask the engineer to review both table privileges and row rules. Supabase’s Row Level Security guide distinguishes those controls and recommends testing allowed and denied operations for client roles. Enabling row-level security, which filters accessible database rows, is only useful when the policies match your intended ownership rules. Privileged service-role access bypasses those policies, so a blocked browser request does not prove that a server function using privileged access checks ownership correctly.

For the removed-member row, agree when removal is supposed to take effect and test that state explicitly. Record whether you used an existing session or signed in again. Check a new CV request as well as the application page. This tests future access; it cannot withdraw information someone already downloaded.

Keep evidence that survives the next change

A forbidden identity might correctly receive a permission error, a deliberate not-found response or an empty result filtered by policy. Read the response as well as its status. A successful denial means the private fields and attachment contents did not arrive through the tested route.

For every row, retain the environment, build, request path, test identity, expected fields, actual result, time and owner. Include a sanitized response excerpt showing the fake marker or the absence of private fields. Exclude session cookies, tokens and real résumés. Keep failed results alongside their retests so a later pass preserves the record of what changed.

If a forbidden identity receives private data, keep real intake closed. Have a qualified engineer fix the enforcement rule, then rerun both the allowed control and the denied cases. Test permitted writes separately with disposable records; this sheet addresses reading candidate data and is not a complete security assessment.

Assign an owner to repeat the relevant rows when roles, authentication, tables, storage or delivery routes change. When the results match the agreed access rules on the intended build, the hiring owner can accept that part of the launch. That completes the read-permission acceptance check for the first real CV.

Related articles

Try Kit for 30 days.

Hiring, security reports, and training in one account, for teams where none of it is a full-time job. Free for 30 days, card required. Cancel before it ends and you pay nothing.

Get started free