The robot software is actually very generalized. CMU's efforts (massively planning after being given the course data) at the 2005 challenge still didn't win them the race. Stanford (the eventual winner)'s implementation for the 2005 challenge was extremely general, and Stanford didn't have to do any data entry after receiving the race data.
I'm not sure where you get the idea that they've been optimized for only this course. The announcement of its general location only happened in September (before that, all teams knew was "somewhere in the southwest USA". Teams will only receive their 'mission' very shortly before the start on Saturday. Yes, it is possible that the winning team will only work well on the DARPA course and will fail elsewhere, but it is much more likely that the winning team's vehicle has fairly generalized sensing and driving skills.
As far as unethical things to do to win $2 million, I'd say it'd be much more likely for a team to be controlling their bot remotely from a substantial distance away. Hopefully DARPA has some countermeasures in place against a scheme like that.