AgentVisor: Defending LLM Agents Against Prompt Injection via Semantic Virtualization | ArxivCSExplorer