#3 doesn't necessarily require breaking RSA, just TLS (via fake certificates or something else). Breaking RSA would be pretty interesting though.
It's safe to assume the NSA can easily do way more than 2 petaflops, and they have an exaflop goal, and that would be enough to run known attacks against DES, factor 1024bit RSA moduli ... and if they can compromise just one root CA (which uses 1024bit RSA) they can issue valid certifications of their own and MITM everyone, and none would be the wiser.
And all of this assuming the NSA relies on publicly known weaknesses in SSL/TLS. The matter of the fact is that they have very smart people with access to unlimited resources researching new vulns and actively exploiting them.