It's a fine line, because after all it is also [part of] the responsibility of the NSA to protect American citizens and businesses from foreign interception. They have two directly competing goals when recommending an encryption scheme: make it as secure as possible to prevent foreign/malicious actors from intercepting it, while they also want to be able to break any encryption scheme they recommend themselves.

To me, this says that they would be more likely to recommend fundamentally sound encryption schemes, but choose parameters where they can have hidden knowledge that lets them undermine it. Stuff like what people think they may have done with the eliptic curve constants, where they [supposedly] have chosen constants that they have extra knowledge about, but the method itself is apparently otherwise quite solid.