Company creates a project and receives a sender ID. Company creates an app, bakes in its sender ID and places the app in the store.
Attacker reverse engineers the app and extracts both the sender ID and the server interface used to receive GCM registration IDs.
Attacker creates his own app, bakes in Company's sender ID and server registration interface, puts app in the store. The attack app basically impersonates Company's real app as far as GCM goes: it registers to receive messages from Company's sender ID and then sends its GCM registration ID to Company's servers just like the "real" app does.
Now Company wants to broadcast some information to all instances of its app. Maybe it's a reminder than an update is available. Is there any way to differentiate the "attack app" (which registered just like the real one) from "real" versions of the Company's app?
The same problem could also have existed with C2DM, which you can sniff the sender email address, instead of project ID for GCM.
C2DM or GCM, should never be used to send sensitive user information (i.e. account name, private information, etc), it's mainly useful for notification, which the real app can use it to perform further actions.
I can't see how useful a notification can be to a 'fake/hack' app, what are they going to do with 'You have new message' notification?