Previous students I worked with used machine learning for state sponsored malware authorship attribution. You can find an abstract here. A data set for benchmarking algorithms on this task can be found here and some code samples here.
Others have worked on learning state representation for Atari games, that can be used to formally verify the behaviour of agents learning to play the game. Learn more here.

If you are looking for a project in the area of safety or security for machine learning, it is best to contact me.