We present an extension of traditional "black box" fuzz testing using a genetic algorithm based upon a Dynamic Markov Model fitness heuristic. This heuristic allows us to "intelligently" guide input selection based upon feedback concerning the "success" of past inputs that have been tried. Unlike many software testing tools, our implementation is strictly based upon binary code and does not require that source code be available. Our evaluation on a Windows server program shows that this approach is superior to random black box fuzzing for increasing code coverage and depth of penetration into program control flow logic. As a result, the technique may be beneficial to the development of future automated vulnerability analysis tools.