A Security Analysis of Long-Horizon Agentic AI Systems: Threats, Evaluation, and Framework Development

This paper presents a structured analysis of security challenges in long-horizon agentic AI systems, detailing existing threats, evaluation methods, and attack propagation mechanisms. It introduces a taxonomy of security threats and a framework for analyzing attack propagation, aimed at guiding future research in the security of agentic AI. This work is significant for practitioners as it provides foundational insights and tools necessary for developing secure long-horizon AI systems.