Deep recommender systems (DRS) are critical for current commercial online service providers, which address the issue of information overload by recommending items that are tailored to the user's interests and preferences. They have unprecedented feature representations effectiveness and the capacity of modeling the non-linear relationships between users and items. Despite their advancements, DRS models, like other deep learning models, employ sophisticated neural network architectures and other vital components that are typically designed and tuned by human experts. This article will give a comprehensive summary of automated machine learning (AutoML) for developing DRS models. We first provide an overview of AutoML for DRS models and the related techniques. Then we discuss the state-of-the-art AutoML approaches that automate the feature selection, feature embeddings, feature interactions, and system design in DRS. Finally, we discuss appealing research directions and summarize the survey.