Agent security is a systems problem

A group of security researchers published a position paper on arXiv arguing that agent security should be treated as a systems problem: the models driving agents should be treated as untrusted components, and security invariants should be enforced at the system level. The paper argues that the community's mainstream approach of improving model robustness is not sufficient on its own and needs to be supplemented with techniques from the systems security field. Drawing on decades of systems security research experience in operating systems, networking, formal methods, and adversarial machine learning, the authors propose a set of core principles to provide a foundation for designing agent systems with predictable guarantees. As evidence, the paper analyzes eleven representative real-world agent attacks, discusses how these systems principles could have prevented the attacks had they been implemented, and points out the research challenges in realizing these principles in agents.