LATIDIA · Ciberseguridad

VirusCascade: Hijacking Collaborative Reflection in LLM-Powered Recommender Agents

arXiv: 2609.38270v1Tipo de anuncio: nuevo Resumen: Avanzando más allá de los modelos tradicionales de puntuación estática, los sistemas de recomendación agéntica impulsados por LLM (LLM-ARS) instancian a los usuarios y elementos como agentes autónomos, cuyo estado semántico

La noticia

arXiv:2609.38270v1 Announce Type: new Abstract: Advancing beyond traditional static scoring models, LLM-powered agentic recommender systems (LLM-ARS) instantiate users and items as autonomous agents, whose semantic states are dynamically refined through a recurrent process known as collaborative reflection. While this mechanism improves recommendation quality, it simultaneously introduces a systemic vulnerability: adversarial evidence injected into a single agent can be rationalised into a legitimate preference narrative, written back into memory, and propagated to other agents through interaction contexts. We term the local rationalisation process reflection laundering, and its system-wide escalation through collaborative reflection collaborative-reflection hijacking. Existing attacks on recommender systems, whether based on interaction-level data poisoning or text-level adversarial perturbations, assume static