They complain about how CBOR is a superset of JSON data types and so some CBOR values (like bignum) might not down-convert to JSON cleanly, and then in the next paragraph they talk about how Ion is a superset of JSON data types including 'arbitrary sized integers'.
Bad doubletalk. Boo. (I have implemented CBOR in a couple languages and like it. Every few months we get to say, "oh look, _another_ binary JSON.")
@LVB, thanks for that. RTFM-ing made me think twice about adopting CBOR or going with Ion. I'll also mention Velocypack (https://github.com/arangodb/velocypack) while here.