Research questionHow can LLM-agent systems prevent safety compromises from propagating across workflow boundaries?Prompt injection, tool misuse, and memory poisoning can appear as distinct attacks even when they exploit related failures at system interfaces. Without a shared view of those interfaces, it is difficult to identify where isolation first breaks or trace the compromise into execution. Latest papersRecent research connected to this question, newest first.Isolation as a First-Class Principle for LLM-Agent System Safety: Concepts, Taxonomy, Challenges and Future DirectionsThe source surveys LLM-agent safety across user-agent, agent-tool, agent-execution, agent-agent, and system-environment boundaries. It provides a conceptual taxonomy of cross-boundary failure paths and open challenges, including isolation-by-construction, rather than evidence from a specific deployment or a tested defense.research paper · Sep 2, 2026