Online social media have become one of the main communication platforms for political discussion. The online ecosystem, however, does not only include human users but has given a space to an increasing number of automated accounts, referred to as bots, extensively used to spread messages and manipulate the narratives others are exposed to. Although social media service providers put increasing efforts to protect their platforms, malicious bot accounts continuously evolve to escape detection. In this work, we monitored the activity of almost 245K accounts engaged in the Twitter political discussion during the last two U.S. voting events. We identified approximately 31K bots and characterized their activity in contrast with humans. We show that, in the 2018 midterms, bots changed the volume and the temporal dynamics of their online activity to better mimic humans and avoid detection. Our findings highlight the mutable nature of bots and illustrate the challenges to forecast their evolution.