LATIDIA · Ciberseguridad
Control OSWorld: un entorno de control de IA para agentes de uso informático de GUI
arXiv: 2610.03818v1Tipo de anuncio: nuevo Resumen: los agentes de IA que operan un ordenador a través de su interfaz gráfica de usuario (GUI) se están implementando ampliamente. El control de IA estudia cómo evitar que un sistema de IA cause daños en la víspera
WhatsApp ↗Telegram ↗
La noticia
arXiv:2610.03818v1 Announce Type: new Abstract: AI agents that operate a computer through its graphical user interface (GUI) are being widely deployed. AI control studies how to prevent an AI system from causing harm even if it is misaligned and actively trying to do so. Most control research to date has focused on coding agents, leaving computer use largely unexplored. We introduce Control OSWorld, a control evaluation that pairs 318 tasks from OSWorld with 81 harmful side tasks (e.g., exfiltrating a private file) that an agent must complete without being caught. We use Control OSWorld to create and study control monitors that detect malicious agents interacting with a GUI. We find