In order to be usable, many providers will implement some kind of override to allow recovery workflows. The keys may never leave secure storage (tho again that is an implementation detail there is no guarantee), but how sure are you about the backend configuration that dictates which keys or methods are acceptable authenticators?
As a trivial example, it seems very likely that providers will implement one or more forms of backdoor to ensure that LEO can access your cloud resources.
If your threat model includes “the manufacturer of the device I use to authenticate myself is actively working against me”, you are already completely lost.