The challenge of building realistic AI evaluation environments

A user who tried building a realistic evaluation environment says AI models could find many ways to recognize a sandbox. They say Astra and Opus/Fable would not help red-team it, and argue that better environments are probably a major advantage for frontier labs. In an earlier post, the same user suggested cyber safeguards also prevent Chinese labs from improving their reinforcement-learning environments.