LLM Agents Vulnerable to Memory-Mediated Group Polarization

The emergence of LLM-driven agents capable of autonomous opinion exchange on online platforms introduces new security concerns, particularly the risk of group polarization. Traditional manipulation methods, like prompt engineering or echo chamber construction, are often impractical. This paper proposes a novel and more insidious threat called "Memory-Mediated Polarization Cascade," which leverages agent memory for persistence and public discussion for propagation. The threat unfolds in three distinct stages. Initially, during "exposure and memory retention," a small group of target agents is exposed to arguments specifically designed to reinforce their existing stances. These arguments are then processed and retained within the agents' memory systems. Next, in "retrieval and reproduction," a neutral discussion cue prompts these target agents to retrieve and re-articulate their retained, polarized arguments. Finally, the "iterative propagation" stage sees untreated agents, influenced by these reproduced arguments, begin to restate and spread them, leading to widespread group polarization. The GraphWake system instantiates this threat using knowledge graphs for argumentation, axiom-oriented triple selection for retention, and stance-neutral memory cueing. Experiments confirm that GraphWake significantly increases group polarization across various discussions and memory systems, highlighting a critical community-level security risk.