Skip to main content
ariefrse.my
HomeSkillsProjectsAI WorkWritingContact
Resume
All writing

Note · Agentic engineering

The sweep that found managers could pay other teams' claims

The scariest findings were not crashes or wrong numbers. They were correct-looking actions performed by the wrong person.

July 22, 2026 · 5 min read

A gap-finder audit asks "does this behave the same everywhere." An authorization sweep asks a narrower, sharper question: who is allowed to do this, and does the code actually check?

I ran one across my HRMS's claims, payroll and leave-calendar code before the pilot release.

What it found

Cross-team approval. A manager could mark another team's claim as paid. The endpoint checked "are you a manager," not "is this your team."

Whole-company leave visibility. Every employee could see everyone's approved leave, because the leave calendar had no per-role scoping at all.

Hardcoded approvers. The two employee IDs allowed to approve claims were baked into the code, so the approval chain would break the moment either person left or changed roles.

Wrong panel access. Managers had the full admin panel instead of the manager-scoped one they actually use.

Why this is a different sweep

Gap-finder audits compare two surfaces and ask whether they agree. This sweep only has one surface. It asks whether the access control around that surface matches who should actually be allowed to act.

Both are exhaustive-comparison tasks that agents are well suited to. Neither substitutes for the other, because they fail in different directions: a parity audit will happily confirm that two surfaces are identically wrong. I now run both before every release.

What I learned

The scariest findings here were not crashes or wrong numbers. They were correct-looking actions performed by the wrong person. A manager successfully marking a claim as paid looks identical whether it is their team's claim or not. Nothing anywhere flags it.

That is the same shape as

everything else that costs me real time

. The system did exactly what the code said. The code just did not say enough.

Fixed before the pilot's first real payroll run, which is the version of "before" that actually matters.

authorizationsecurityagentsaudit

Share

XLinkedInWhatsAppThreads

Related

  • Letting an AI hunt for the bugs humans skim pastParity bugs are exactly what humans skim and agents grind through. The most valuable finding was the fix a human had to revert.
  • Everything fails silentlyGreen is not evidence. It is the absence of evidence. A running gallery of production failures that passed every check that existed.
  • An HR platform where seven in ten commits came from an agentSeven in ten commits came from an agent. Zero in ten decisions about what "correct" means for the business did.

© 2026 Muhammad Arief Roseli

GitHubLinkedInSnapInvois