Isolation as a First-Class Principle for LLM-Agent Safety: A Boundary-Centric Survey

An arXiv survey argues isolation should be a first-class principle for LLM-agent system safety, using five boundaries—user-agent, agent-tool, agent-execution, agent-agent, system-environment—to unify seemingly scattered attacks like prompt injection, tool misuse, and memory poisoning.
Bottom line
Once an LLM agent becomes the 'brain' of a system, safety is no longer just input-output alignment—it's about system behavior and real execution outcomes. This survey unifies a fragmented attack surface under one principle: isolation.
Five boundaries
The authors propose a boundary-centric taxonomy where isolation lives at five interfaces:
- user ↔ agent
- agent ↔ tool
- agent ↔ execution
- agent ↔ agent
- system ↔ environment
Why it's useful
- Explains why prompt injection, tool misuse, and memory poisoning often share a root cause—a loss of isolation somewhere, propagating through the workflow.
- Helps locate where isolation first breaks, how compromise spreads across boundaries, and which defense fits each interface.
- Proposes an 'isolation-by-construction' research agenda.
For teams building agent systems and their security, this boundary framework is directly usable for threat modeling and defense selection.