Second order methods are attracted to saddle points in high dimensional spaces. The math and practice of optimizing these surfaces has a lot of nuances like this so much of the stuff you learn in your convex optimization class doesn't apply too well.