New AI designs (world design + architectural priors + training/education system) should be tested first in the safest virtual worlds: which in simplification are simply low tech worlds without computer technology. Design combinations that work well in safe low-tech sandboxes are promoted to less safe high-tech VR worlds, and then finally the real world.

A key principle of a secure code sandbox is that the code you are testing should not be aware that it is in a sandbox.

So you're saying that I'm secretly an AI being trained to be friendly for a more advanced world? ;)