As AI models become more capable, understanding how they’re tested and contained is becoming increasingly important.

source