Like many other animals, humans direct their gaze to selectively sample the visual space based on task demands. Visual search, the process of locating an item among several visually presented objects, is a key paradigm in studying visual attention. While much is known about the brain networks underlying visual search, our understanding of the neural computations driving this behavior is limited. Simulating this behavior in-silico can provide a tool for developing and testing hypotheses about these neural computations. To address this gap, we trained an image-computable artificial neural network to perform overt visual search in natural scenes. Our model consists of an approximate retinal transformation, a convolutional network mimicking the neural computations along the ventral visual pathway, and a recurrent neural network or transformer model of the fronto-parietal attentional network. After training, the model demonstrated strong generalization in search performance to unseen images while exhibiting high behavioral consistency with human subjects, surpassing current visual search models without being trained on eye-tracking data. A representation of the retinocentric and allocentric priority maps, akin to those described in the primate brain emerged in the model’s latent space. These representations persisted in time and were updated with each saccade. Further analysis of the model’s latent space revealed some predictions about the neural representations in the primate brain during visual search: 1) Cue information is consistently encoded in the same neural subspace; 2) Retinocentric priority map is consistently encoded in a continuous representation, i.e. the priority at locations in closer proximity in the visual field are encoded in neural subspaces that are more aligned. Our model provides a computational framework for further studying the neural circuits underlying visual search in the primate’s fronto-parietal cortical network.