Roshan Trivedi newsletter icon: the letters RT on a blue to violet gradient

Roshan Trivedi

Archives
Log in
Subscribe
November 2, 2026

I tested 27 machine identity cases. My guesses were wrong in two places.

Hello,

The last issue was about AI governance. This one is about machine identities: service accounts, API keys, CI tokens, break-glass accounts and AI tools that quietly hold their own logins.

Most companies have many more of these than people. A mapping sheet says a control exists. I wanted to see where it stops.

How I tested it

Before any code, I wrote down what I expected and committed the files. Then I wrote small models of each control and ran 27 cases with the control off and on. Then I switched each control off by itself to see whether anything noticed. Everything is made-up data, and a match proves little because I wrote both the guesses and the code. The misses are the useful part.

Where I was wrong

My secret scan caught 6 of 8 fake secrets. I guessed it would miss one. It missed two: a secret split across two lines, and a base64-encoded one.

I also guessed every control would show up when switched off alone. Two did not. A 10-minute token lifetime changed nothing once the token was bound to the CI runner and revoked at job end. A vendor-breach revoke changed nothing once an app allowlist refused the grant. Each attack was already stopped by something else. That is layering, and it means a one-at-a-time test can say "nothing happened" about a control that matters in a different scenario.

Where controls fell short, as expected

  • A made-up ticket number passed the break-glass review because nothing checks it against the ticket system.

  • A false justification for using break glass was accepted.

  • An allowlisted app that is later breached keeps working until someone revokes it.

  • A bound CI token still works from inside the runner during the job.

Three of the 27 cases came from 2026 incident reports, including a third-party AI app that an employee had linked with broad permissions. Reports on that incident differ in detail, so I linked the sources on the page.

Try it.

The page has an animated workflow for every case and a small lab you can run in your browser: Break a vendor OAuth grant, steal a CI token, and paste fake secrets into the scan. Nothing leaves your browser.

roshantrivedi.co.in/nhi-lab

One question for you

If you switch one of your controls off, does anything tell you? If not, is another control covering it, or nothing?

Reply and tell me what you would test first. If you can think of a variant that breaks a control, I will add it and keep the original on the page.

Roshan

Don't miss what's next. Subscribe to Roshan Trivedi:
← Newer I never said what I was counting Older → Welcome — here's what to expect
roshantrivedi.co.in
GitHub
LinkedIn
Powered by Buttondown, the easiest way to start and grow your newsletter.