There are so many statements about the risks of using AI agents, but unfortunately not many well-structured ways to prevent privacy leakage to public resources. It's certainly challenging to find a compromise between a convenient user experience and data privacy. We believe every AI agent user should ask themselves the following questions: What can AI agents actually do? What do you want them to do? What data must be restricted? We found an interesting way for macOS users to get some answers using Apple's built-in features. Every macOS application or executable binary can be sandboxed using built-in policies, described in a sandbox rules file. The rules language is simple and intuitive, e.g. "allow file-read*", "allow file-write*". The idea is to prohibit all outbound connections and file/directory access, then check the logs with log stream | egrep Sandbox. As a result, you can monitor what the agentic software tried to access.