We present a straightforward and efficient way to estimate dynamics models for unstable robotic systems. Specifically, we show how to exploit the differentiability of Gaussian processes to create a state-dependent linearized approximation of the true continuous dynamics. Our approach is compatible with most Gaussian process approaches for system identification, and can learn an accurate model using modest amounts of training data. We validate our approach by iteratively learning the system dynamics of an unstable system such as a 9-D segway (using only one minute of data) and we show that the resulting controller is robust to unmodelled dynamics and disturbances, while state-of-the-art control methods based on nominal models can fail under small perturbations.