What is the best way to sandbox LLMs and packages in general, while being able to work on data from outside sandbox (get data in and out easily)?

There is also the need for data sanitation, because the attacker could distribute compromised files through user’s data which will later be run and compromise the host.