IP hijack detection is an important security challenge. In this paper we introduce a novel approach for BGP hijack detection using deep learning. Similar to natural language processing (NLP) models, we show that by using BGP route announcements as sentences, we can embed each AS number (ASN) to a vector that represents its latent characteristics. In order to solve this supervised learning problem, we use these vectors as an input to a recurrent neural network and achieve an excellent result: an accuracy of 99.99% for BGP hijack detection with 0.00% false alarm. We test our method on 48 past hijack events between the years 2008 and 2018 and detect 32 of them, and in particular, all the events within two years from our training data.