Sandboxing Agency: Isolation Protocols for Third-Party Tool Use

ABSTRACT

Current agents often utilize third-party tools (APIs, web browsers, databases) with full authority, creating a 'Tools-as-Attack-Vector' problem. We introduce 'Agency Sandboxing,' a software engineering protocol that enforces strict isolation between an agent's reasoning core and its execution capabilities. By introducing an intermediary 'Gatekeeper Agent' that audits every tool call against a predefined safety policy, we can prevent 'privileged escalation' where an agent is tricked into using a tool outside its authorized scope. Our results show that sandboxing reduces the risk of malicious data exfiltration by 92% with a negligible impact on task completion time.