Memory-Mediated Polarization Cascade Threat in LLM Agent Communities

Attackers can induce group polarization in autonomous agent communities by exploiting agent memory as a persistence channel. This three-stage attack uses memory retention of biased arguments and subsequent retrieval during neutral discussions to drive iterative propagation of polarized stances.