"Why shouldn't we assume the NSA has subverted the CA system?"

Broadly speaking, we should. And that's why there's been a lot of work lately in doing things that either build on or bypass the CA system. For instance, even if you get a valid CA certificate for a site you want to spoof, it still has to be a different certificate, so there's more work on certificate pinning. Sure, that doesn't help if the spoof cert is up the first time a given person visits a site, but it still raises the bar for successfully attacking a site surreptitiously, because anyone who hits the site post-pin will get a big warning, which can be picked up with other things. Now the attack has to be even better targeted.

And there's other work going on lately, and while it may not be immediately visible to the general public because this all takes time, it really seems to me this has all gotten a lot more vigorous over the last year. I don't even think it's all Snowden per se... it seems to me generalized security consciousness has gone up lately, like we passed a critical threshold of people's work feeding back on other people's work and the community as a whole has moved to a new level of effectiveness. It'll still be a while before we really see the changes but I think some changes are coming. (Not Utopia. We've still got a long ways to go. But some general improvements.)

Then you must replace all traffic to that site successfully. The keys for the spoofed and the legitimate certificate will be different. With PFS only traffic with the negotiated key can be decrypted by the browser. Without PFS, the legitimate server will not understand your requests. Quantum Insert is not successful for each request, as stated by the article. Therefore I would assume that at least a few packages of the original server would slip through which then cannot be decrypted or contain an error message of the server. Especially targeted administrators would notice very quickly, I assume (although browsers might suppress those packages silently). Therefore this would work only with a full man-in-the-middle.

I think in general https would work against this particular attack, if you don't accept sites with mixed content (I am not aware of any browser plugin which does suppress plain content on mixed sites).

It has been. If I recall correctly it was revealed even years before Manning and Wikileaks.

As a matter of observation, it is kind of amazing how short of a memory even the internet has. It seem to be a knowledge management issue, i.e., new and evolving information is not contextualized or even related to pertinent past information.

Or they crack or compromise the keys that they need on a per operation bases. Even assuming ubiquitous SSL (or something more decentralized), the keys of a few advertising or web analytics sites would get a lot of coverage.

Creative use of advertising platforms is used you could easily narrow targets down and guarantee that you know exactly what the content is.

Anyways, this is just one gadget.

Because that subversion can be relatively easily detected, and then shouted to the world.