The good news about this particular bug is that carry propagation bugs are well-studied (they seem to be the textbook case of a bignum math bug), and if OpenSSL says they've scoped down the places this bug impacts, they're probably right.
I don't know if it mitigates your existential dread, but carry propagation bugs in multiplication routines are not in fact all that rare.
True. But they tend to be well-studied by computer security people, not by mathematicians; and most computer security people won't have the necessary number theory / group theory background.
This is precisely the class of bugs for which proper assessment requires mathematicians who study computer security... aka. NSA employees.
In addition, I think it could produce a wrong symmetric key, as a result of a bad computation e.g. in the Diffie-Hellman protocol. So you could have insecure symmetric (e.g., AES) keys.
Granted, finding a concrete example would probably be very difficult, considering that replacing modexp by rand() is only going to pass N as a suspected prime with probability 1/N; but the question was can it, not will anyone ever find an example.
His post is shot in black and white, as he is smoking while sitting in an cafe on an otherwise abandoned European street, when the desultory wind blows in a creased newspaper with the headline "MODULAR MATH FOUND NOT TO WORK", and he turns to look at the camera with a slowly developing look of pure horror that the camera person spends a full 5 minutes in exploring every last nuance of. Just over two minutes in, a violinist wearing a tutu wanders in, plays some 12-tone music, delivers a poorly subtitled rumination on the ultimate futility of life and its connection to prime number elliptic curves, and then leaves, for no apparent reason.
It's moving stuff, for certain audiences.
I haven't looked at the code in question, but I'd be willing to bet that the NSA has figured out a way to exploit the broken multiplication which is completely unlike what the OpenSSL authors were considering.
This one is absolutely terrifying. There are common ways that TLS stacks break: Generally speaking, bugs in parsing certificates, bugs in the protocol handling, and side channel attacks on the cryptography. These are risks which we aim to work around -- just like infections are a known complication of surgery, and antibiotic prophylaxis is recommended in some cases.
A bug which results in multiplication producing the wrong answers is in a completely different category. This is a "never event", akin to a surgical scalpel being left inside a patient after surgery or the wrong leg being amputated.