Isolation as a First-Class Principle for LLM-Agent System Safety: Concepts, Taxonomy, Challenges and Future Directions

Huihao Jing, Wenbin Hu, Shaojin Chen +10

The paper surveys how isolating components such as user inputs, tools, execution, inter‑agent communication, and environment can improve safety of LLM‑agent systems, presenting a b…