Isolation Key for LLM Agent System Safety.

As large language model (LLM) agents become the "brain" of complex systems, their safety concerns extend beyond simple input-output content alignment to encompass system behavior and real-world execution outcomes. The current fragmented understanding of agent failures, such as prompt injection or tool misuse, makes it difficult to identify common structural causes and how these issues spread within an agent's workflow. This comprehensive survey introduces "isolation" as a primary principle for ensuring LLM agent system safety. Isolation refers to the deliberate separation of various components, including user inputs, tool access, execution channels, inter-agent communication, and environmental context. The authors propose a taxonomy centered on five critical boundaries: user-agent, agent-tool, agent-execution, agent-agent, and system-environment. This framework helps pinpoint where isolation failures originate, how compromises propagate across these interfaces, and which defensive strategies are most relevant at each point, ultimately guiding future research towards "isolation-by-construction" for safer agent systems.