Your description of secure communication (Tor, anonymous XMPP, etc.) is totally accurate and a great explanation for those who may not be quite as familiar with security and OPSEC.
Unfortunately, with great enough amounts of metadata and computing power, this could theoretically be correlated as well. DNS requests or Tor connections correlated with encrypted messages timed between two individuals seems like it would be too hard to make meaningful, but it's not impossible. There's precedence for this when the FBI suspected Jeremy Hammond of being a certain identity on IRC, and correlated his sign-ons with connections to Tor.
You're absolutely, 100% correct, though, that this data (even correlated), must be part of some sort of targeted surveillance. I wasn't trying to counter the dragnet argument as much as provide clarity on what these security measures would or would not do.
Thanks for providing even more clarity in the areas I didn't address :)
With a total record of the entire Internet (global passive adversary), you could definitely defeat anything using Tor, but I wonder if you could make it significantly harder still by randomly generating XMPP accounts every time. There's a potentially infinite space, and you can do XMPP registration in-band.
Of course, you could also just run XMPP servers locally as chat endpoints, and never have subpoena-able records. These could be Tor hidden services, which would give you the added bonus of global reachability. I think this eliminates most correlation attacks you could do; all you'd see on Alice and Bob's endpoints would be Tor circuits.
All this will make for a really interesting Wire reboot, if/when that happens.
Tor won't help at all if all of the long-distance network traffic in the country is being mirrored (as it has been in the USA for most of a decade).
Corollary: The NSA knows exactly who runs The Silk Road. Stopping drug trafficking is obviously not as high a priority to them as not letting potentially kinetic adversaries know that Tor provides no anonymity to someone who can (and does) monitor _all_ network traffic.
The article is in response to a dragnet surveillance program, where everyone's communications are watched and presumably datamined. It's very easy to do this, because nothing is encrypted, and everyone uses services that expose metadata (like who is IM'ing who).
Your comment is entirely true. However, it presents an adversary that doesn't want dragnet, but targeted surveillance. It assumes that Bob will be immediately arrested if his communications become encrypted.
This is not the threat model that we're faced with now. Let's say you and Bob communicate using accounts you've made on random XMPP servers using Tor, and all the messages are encrypted with OTR. Both servers are in the US, and the NSA's metadata database shows [email protected] sending lots of ciphertext to [email protected].
This is "NSA-proof" in that the NSA would not know to link [email protected] with you using their existing systems. They would have to drastically escalate the cost of their surveillance program with respect to you and Bob to figure out what you're talking about. Unless you really are a political dissident, conspiracy theorist who accidentally discovered the UN's black helicopter program, or radical Islamist, you are now out of the surveillance dragnet.
That is to say, unless the threat model changes, using privacy-enhancing technology will keep your data safe from PRISM and similar dragnet programs.