In the last years, tremendous advances have been achieved in the development of neuroprostheses that allow patients with advanced motor deficits to recover a certain degree of independence. The concept behind these motor brain-machine interfaces is that patients use their motor cortical activity to control ever more complex robotic arms to interact with their environment. In contrast, in spite of the large range of potential applications in the field of cognitive disabilities, only smaller advances have been achieved in the field of cognitive brain-machine interfaces, due to specific methodological challenges. Here I will demonstrate real-time tracking of covert spatial attention from the activity of bilateral prefrontal dense neuronal recordings in the non-human primate, and show that this allows us to predict behavior. I will then discuss how this cognitive information is multiplexed in the brain with other sources of information and how capturing this multiplexing further benefits the real-time decoding of attention. Last, I will provide evidence that voluntary control onto the decoded attentional spotlight, thanks to neurofeedback, improves spatial attention performance through a massive though constrained reorganization of cortical information. Overall, I will thus demonstrate the high potential of cognitive brain-machine interfaces.