Who says the agent is allowed to do this?

The useful question is not “was the model aligned?” It is: was this action allowed, under the rules you have, at the moment it tried to run?

Nexus checks that before the agent reaches the system. Risky steps go to a named person, with a written reason. You keep a record of yes and no.

The same check works anywhere an agent can do something you cannot undo in a chat window. Your core systems stay.