Nothing it does will fail a permission check. The only way to catch it is to know what it normally does, which means knowing the estate it acts on — so the register that says what you have is the same system that says what is behaving strangely.
An agent is frozen at the gate it acts through, everywhere it acts. A person, a schedule or a device key is denied at the junction, on devices enrolled in enforce. An integration or a collector is raised to an owner and nothing is stopped — those levers are not built, and the product refuses to pretend otherwise rather than showing an actor as held while it keeps working.
It catches a campaign over an hour rather than a single action in the instant. Detection runs on a sweep every ten minutes across the preceding hour, so a finding reaches the gate on the next sweep. It also watches writes and dispatches only — reads are not recorded in the streams it is built on.
No. It ships disarmed and you arm it one kind of account at a time. On a disarmed organisation a finding records what would have happened rather than doing it, and a kind of account with no containment lever behind it cannot be armed at all — the server refuses, not just the screen.