Saccades allow us to visually explore our environment. Like other goal-directed movements, their accuracy is permanently controlled by adaptation mechanisms that, in the laboratory, can be induced by systematic displacement of the "real" visual target during the saccade. However, in an anti-saccade (AS) task, the target is "virtual" because gaze has to be shifted away from the "real" visual target toward its mentally defined mirror position. Here, we investigated whether the brain can adapt movements aimed at a virtual target by trying, for the first time, to adapt AS. Healthy human volunteers produced leftward AS during three different exposure phases in which a visual target provided feedback after the AS. In the adaptation condition, the feedback target appeared after completion of the AS response at a location shifted outward from final eye position (immediate non-veridical feedback). In the two control conditions, adaptation was prevented by delaying (800 ms) the shifted feedback target (delayed-shift) or by providing an immediate but veridical feedback at the mirror position of the visual target (no-shift). Results revealed a significant increase of AS gain only in the adaptation condition. Moreover, testing pro-saccades (PS) before and after exposure revealed a significant increase of leftward PS gain in the adaptation condition. This transfer of adaptation supports the hypotheses of a motor level of AS adaptation and of a visual level of AS vector inversion. Together with data from the literature, these results also provide new insights into adaptation and planning mechanisms for AS and for other subtypes of voluntary saccades.