If you control a significant percentage which is less than 50%, you can attempt to do this but it is both uncertain to work (on any given attempt) and costly if it fails. Note that you can take many bites at the apple if you're willing to suffer the opportunity cost, so if you have a sufficiently profitable transaction to cheat on, you'll come out ahead.
The most vulnerable players in the Bitcoin system currently are gambling sites, but in principle with sufficient hash power you can do this to anybody using Bitcoin. Gambling sites are particularly at risk because their intended behavior is swapping bitcoins for bitcoins very quickly, and if you can doublespend, you can swap bitcoins for bitcoins but then say "You know those bitcoins I sent you? Psych, I didn't have them, even though you thought I did." after you've been told that you lost a bet. If you win the bet, you simply don't rewrite history to invalidate your bet. Repeat as desired.
This is the planned and anticipated vulnerability in Bitcoin. Great news: it's rarely the planned and anticipated vulnerability that kills a system.
+ Double spend attacks are not the only attacks you can envision. For example, if you control 51% of hashing power for a month, you're capable of essentially invalidating all transactions globally in the past month, at a time of your choosing. Like, if an unrelated party Bob had been paid in bitcoin by his employer two weeks earlier, and Bob then attempted to transfer bitcoins to Mt. Gox to change into dollars, Bob's bitcoin client would suddenly tell him "Dude, you might remember that salary payment, but it never actually happened." and Mt. Gox would say "Umm, you don't have bitcoin to transfer in, what are you talking about?"