You can have the exact same problem with just different versions or even configurations.
I think the idea is that the protocol is resilient to differences in implementation. A bad actor, could for all intents and purposes, manipulate the implementation to their advantage. An implementation, reference or not, is merely a convenience for users who do not want to write their own clients (which is essential for adoption).
The problem with multiple implementations is that if there is a tiny difference in behavior between them, you can only find out after damage has been done to the network via a split/fork.