A friend of mine's working on a bunch of DARPA projects, and his path was:
1. Attend a top liberal-arts college (CS, Amherst)
2. Get master's degree from top research university (HCI, Carnegie Mellon)
3. Work for General Dynamics
4. Stick around long enough to get promoted, kick butt, etc.
5. Profit!!!
BTW, he said that chances are, whatever the next big thing is will come out of DARPA. His bet was on pervasive computing - DARPA's apparently doing a lot with computers in clothes, computers in canteens, computers in backpacks, computers in visors, basically computers everywhere. The logic being that most of their troops are getting deployed to inhospitable environments (Iraq, Afghanistan, etc.) and if they can monitor environmental conditions for things like excessive heat, sandstorms, laser sniper sights, explosive residues, they can save lots of soldiers' lives. And of course, if it trickles down to consumer uses it could completely change how people use computers.
1) A lot of large research companies get DARPA research grants. General Dynamics (like nostrademons said), HRL, Boeing, etc. etc. You can get a job with them, and if the environment doesn't feel soul-crushing to you, you are golden.
2) A lot of CS research at Universities is funded by DARPA. If you are still in school, or if you are out but still have close relationships with your professors, you can always find out what kind of projects they have. DARPA projects are especially nice because they are large in scope and there's always a need for extra people. Don't expect to be paid much though.
I chose route 2 - after my startup failed I wanted to do something that paid, was not as intense, and had nothing to do with Web 2.0 and Social Networking and all that, so I just e-mailed a professor.
I'm not sure DARPA is that hot these days. Having worked on a DARPA project and talked to (and gone drinking with) a lot of people in academia and industry that work on DARPA projects, it seems they're suffering from focusing too hard on yearly "deliverables".
In my experience, that makes it difficult to get anything too interesting done as people (due to budgetary constraints, particularly at the industrial research labs) focusing on getting the minimum deliverables done, instead of trying new things. The downside here is pretty obvious: if a DARPA program manager could specify 3 years in advance what should be achieved, then it isn't really research, more like engineering.
I guess that's why most savvy researchers only write DARPA/NSF proposals for things they've already done (but not published on); gives them money to work on the _next_ big thing :-)
DARPA has all of the amazing stuff coming out of it...