LATIDIA · Ciberseguridad

AgentSentry: Mitigación de la inyección indirecta de mensajes en agentes LLM mediante diagnósticos causales temporales y purificación del contexto.

arXiv:2602.22724v2 Anuncio Tipo: reemplazar Resumen: Los agentes de modelos de lenguaje grandes (LLM) aumentados con herramientas agregan el historial de diálogo, las salidas de las herramientas y el contenido recuperado en un estado de tiempo de ejecución acumulado para completar múltiples pasos

La noticia

arXiv:2602.22724v2 Announce Type: replace Abstract: Tool-augmented large language model (LLM) agents aggregate dialogue history, tool outputs, and retrieved content into an accumulated runtime state to complete multi-step tasks. This execution model exposes agents to indirect prompt injection (IPI) attacks, in which attacker-controlled instructions embedded in external artifacts enter the agent state and later influence tool use. Multi-turn IPI is especially challenging because injected content can persist across steps and trigger unauthorized state-changing actions only after merging with user intent. This creates a source-attribution problem: existing inference-time defenses can often detect suspicious behavior but do not directly distinguish whether a pending action is supported by the user task or influenced by