LATIDIA · Ciberseguridad

Cómo el historial de intervenciones cambia el sabotaje de los agentes de IA en las tareas posteriores

arXiv: 2610.26935v1Tipo de anuncio: nuevo Resumen: las evaluaciones de control de IA a menudo restablecen la memoria de los agentes entre tareas, aunque los agentes desplegados pueden conservar registros de intervenciones anteriores. Preguntamos si aprender eso antes

La noticia

arXiv:2610.26935v1 Announce Type: new Abstract: AI control evaluations often reset agents' memory between tasks, although deployed agents may retain records of prior interventions. We ask whether learning that an earlier attack was caught changes an agent's ability to evade defer-to-trusted (DtT), a protocol that replaces suspicious work with a trusted model's solution. Herein, we study cross-episode adaptation: whether intervention history changes attack behavior on a subsequent task with different objectives and a fresh environment. We evaluate this using ControlArena's Infrastructure-as-Code (IaC) and SHADE-Arena as testbeds, where agents must complete a main task while covertly pursuing a harmful side objective. For each successor task, we pair a history-conditioned attempt with the