[e7h4nz]

In this problem domain, I believe humanity is still in a very early stage. What we can do is treat the agent and its operating environment as a "black box" and audit all incoming and outgoing network request traffic.

This approach is similar to DLP (Data leak prevention) strategies in enterprise-level security. Although we cannot guarantee that every single network request is secure, we can probabilistically improve safety by adjust network defense rules and conducting post-event audits on traffic flow