Agent Sandboxing

Security Sandbox for AI Agents

Definition

A restricted execution environment that limits what an AI agent can access or modify on a system, preventing unauthorized file operations, system calls, or data access even if the agent is compromised or receives malicious instructions

Examples in the Wild

  • Example 1:Phylax preventing agents from reading or deleting files
  • Example 2:Container-based agent isolation
  • Example 3:Intent-based access control for agent operations