We investigate the effect of externally manipulated hub users (influencers) on public opinion in online social media, and discuss how this triggers polarization through repulsion from opposing views. To analyze this, we propose a multiagent-based opinion dynamics model that reflects the asymmetric influence structure of social media platforms. By integrating established models of opinion dynamics (Friedkin-Johnsen, bounded confidence, and social reinforcement), we reveal an emergent backfire effect—a counter-intuitive polarization mechanism. The model incorporates social influence, nonstatic bounded confidence, social rewards through posting, follow/unfollowing, and features specific to social media environments. Experiments on scale-free networks show that steering influencers with moderate opinions toward extremes provokes repulsive responses from users with opposing views, depending on the influencers’ network position. These findings highlight how targeted manipulation of well-connected, neutrally positioned influencers can unintentionally intensify polarization, highlighting the structural role of influencers in shaping opinion polarization.