Cyber-physical systems (CPSs) play a critical role in automating public infrastructure and thus attract wide range of attacks. Assessing the effectiveness of defense mechanisms is challenging as realistic sets of attacks to test them against are not always available. In this short paper, we briefly describe smart fuzzing, an automated, machine learning guided technique for systematically producing test suites of CPS network attacks. Our approach uses predictive machine learning models and meta-heuristic search algorithms to guide the fuzzing of actuators so as to drive the CPS into different unsafe physical states. The approach has been proven effective on two real-world CPS testbeds.