Sandboxing is helpful for "inbound" attacks from scripts/skills toward the host system, but does nothing for "outbound" attacks from the script/skill toward networks -- which includes both the internet, the intranet (including WANs), SDNs, the LAN, and even localhost (including other user accounts on localhost).

Primarily the way to address this is with vetted skills, but network lockdown, including lockdown per skill/agent, should also be used.