Pattern
An agent able to use tools or act on a system should not share an environment with sensitive personal applications and data by default.
Practice
Run agentic AI in a separate machine, account, virtual environment, or tightly scoped workspace that cannot reach personal data it does not need.
Start small
Check what one agent can read or execute and remove one permission it does not need.
Watch out
Sandboxing reduces risk but is not a guarantee; treat broad or unattended access as a meaningful attack surface.