The point about a valid token not being meant for your app is the part I would underline. It is easy to treat a successful login as proof that the request is safe, and then skip the checks that actually decide what the user may do. I also liked the reminder to validate permissions on the backend on every request, not only in the interface. Hiding a button is not a security control. Reading every change an AI coding agent makes, the way you describe, is the right habit here, because these are exactly the gaps that look finished at first glance.