Mapping points to the curve, multiplying points, point addition, there is too much math stuff to mess up on.
Just don't do it. If you read one thing tptacek says, read the last paragraph.
Well, ultimately, somebody has to.
Its probably better to have more qualified people working on (analyzing and validating, particularly) a smaller number of production implementations than fewer people per implementation doing that with a larger number of production implementations.
Encryption is an area where doing one, very precisely defined, task correctly is critically important; the considerations in that domain are different than in many other domains of software.
Alex Balducci in our office actually got this attack working from the paper and walked us through the code --- the lattice reduction steps take 8 hours to run, and are followed by an IFFT-aided search that I would have zero chance of getting right and so never would have bothered waiting 8 hours to try, but he did it anyways.
Maybe I can get him to write it up.
Moral of this story by the way: hire people smarter than you are, and give them semi-unreasonable problems to work on.
Same thing with timing attacks, til I learned how to code one for myself.
It's funny how very very careful you have to be.
The problem with NIST Dual_EC_DRBG is simpler than the article makes it sounds. A good mental model for Dual_EC is that it's a CSRPNG specification with a public key baked into it (in this case, an ECC public key) --- but no private key. The "backdoor" in Dual_EC is the notion that NSA --- err, Clyde Frog --- who is confirmed to have generated Dual_EC, holds the private key and can reconstruct the internal state of the CSPRNG using it. I think this problem is simple enough that we may do a crypto challenge on a toy model of Dual_EC.
Nobody in the real world really uses Dual_EC, but that may not always have been historically true; the circumstantial evidence about it is damning.
The NIST ECC specifications are in general now totally discredited. If you want to see where the state of the art is on ECC, check out http://safecurves.cr.yp.to/.
You should never, ever, never, nevern, nervenvarn build your own production ECC code. ECC is particularly tricky to get right. But if you want to play with the concepts, a great place to start is the Explicit Formulas Database at http://www.hyperelliptic.org/EFD/; the fast routines for point multiplication are mercifully complicated, so copying them from the EFD is a fine way to start, instead of working them out from first principles.