Systems Security Foundations for Agentic Computing

A group of security researchers published a paper on arXiv analyzing the security problems of agentic AI systems from the perspective of classical cybersecurity. The paper points out that agents, because they continuously interact with third-party servers, face risks such as data exfiltration caused by prompt injection, while most existing literature starts from an AI perspective and lacks a systems security perspective and guarantees. The authors take the end-to-end security properties of the entire system rather than isolated models as the object of study, identify short-term and long-term research problems in agentic AI security, and distill these ideas for AI practitioners. The paper also covers 11 real-world attack case studies against agent systems and proposes a series of new research problems specific to this field.