LATIDIA · Ciberseguridad

Cuando la experiencia se convierte en una puerta trasera: envenenar a los agentes de LLM autoevolucionados a través de la optimización de tareas conscientes de fallas

arXiv: 2611.10542v1Tipo de anuncio: nuevo Resumen: los agentes del modelo de lenguaje grande (LLM) se mejoran cada vez más a través de la autoevolución, destilando la experiencia de ejecución en habilidades persistentes y reutilizables. Sin embargo, este pe

La noticia

arXiv:2611.10542v1 Announce Type: new Abstract: Large language model (LLM) agents increasingly improve themselves through self-evolution, distilling execution experience into persistent, reusable skills. However, this persistence creates a new attack surface: malicious rules implanted into skills can compromise subsequent agent behavior. Existing skill backdoor attacks typically assume direct manipulation of the victim's skills, which is impractical when skills are constructed internally by a trusted evolver. In this paper, we study a novel problem of task-level backdoor attacks, where the attacker can only inject poisoned evolution tasks, inducing the agent to implant backdoors into its own skills. Since the attacker controls neither skill induction nor downstream execution, a natural approach is to