We introduce Neural Programming (NP), a novel paradigm for writing adaptive controllers for Cyber-Physical Systems (CPSs). In NP, if and while statements, whose discontinuity is responsible for frailness in CPS design and implementation, are replaced with their smooth (probabilistic) neural nif and nwhile counterparts. This allows one to write robust and adaptive CPS controllers as dynamic neural networks (DNN). Moreover, with NP, one can relate the thresholds occurring in soft decisions with a Gaussian Bayesian network (GBN). We provide a technique for learning these GBNs using available domain knowledge. We demonstrate the utility of NP on three case studies: an adaptive controller for the parallel parking of a Pioneer rover; the neural circuit for tap withdrawal in C. elegans; and a neural-circuit encoding of parallel parking which corresponds to a proportional controller. To the best of our knowledge, NP is the first programming paradigm linking neural networks (artificial or biological) to programs in a way that explicitly highlights a program's neural-network structure.