tl,dr: the first thing one does in this field is to sanitize the input.

The thing is, they add vulnerabilities in a number of ways. They can do it directly with code. They can do it indirectly with standards hard to code correctly w/out vulnerabilities or side channels. There's lots of options. Whatever they do will usually look like a helpful contribution or useful requirement that went wrong in a way that leads to an attack. The better ones are those that look like common or inevitable errors. That's because obvious backdoors make folks run away from a project or supplier maybe forever on top of question who put it there. So, it's usually these flaws that look like obvious errors that still get the job done with everyone around defending the person that put them there.

And I'm not saying it was an NSA job. I have no idea. They've been doing too good of a job on most things for me to know. Could've been an accident. Even probability supports it's an accident just like it did all the times it was a subversion. At $200+ mil a year budget for backdoors/hacking, you bet there were a lot of accidents that, in non-TS version, had nothing to do with the NSA. ;)

Edit: There's lots of questionable things in this article. My favorite is this:

"And this group are the best of the best of the best."

The OpenBSD team doing LibreSSL had all kinds of summaries, live updates, and even presentations of what they found. It was about as far from that quote as you could imagine. Although my memory sucks, I think at one point they said there was even code that checked to see if endianness changed while it was operating. They at least had that covered. There were so many oddities about that codebase.