ANU Study Reveals LLM Social Networks Vulnerable to Opinion Polarization Attacks

A new study from Australian National University (ANU) provides the first systematic analysis of how adversaries can amplify opinion polarization in LLM-based social networks. Unlike traditional mathematical models that rely on simplified assumptions, the researchers built a simulation where LLM agents with diverse personas interact through natural language posts and update opinions contextually. This approach captures realistic adversarial strategies—such as persuasive or manipulative messaging—that classic models cannot represent. The study found that even an adversary with a limited manipulation budget can considerably increase polarization across the network.

To counter these attacks, the team evaluated two classes of defense mechanisms: reactive mitigations, which assign specific users to actively counter manipulation, and proactive interventions, which increase general resistance without targeting specific users. While both methods reduced the impact of adversarial attacks, neither restored the network to its baseline polarization state. These results suggest that current mitigation strategies are insufficient to fully protect LLM-based social systems from manipulation, underlining the potential risks as AI-powered social platforms become more prevalent. The paper includes 14 pages and 7 figures, with implications for platform designers and policymakers.