We present a new approach to exact probabilistic inferene in hybrid Bayesian networks—networks that contain both continuous and discrete variables. The method is the first exact inference method that allows for discrete variables to have continuous variables as parents and non-linear relationships among continuous variables. The family of distributions to which our method can be applied include those defined by a CART network—Bayesian networks in which the local structure associated with each variable is either a classification or regression tree.