Comment by embedding-shape

I'm fairly sure most "safety" people consider "large scale automated crime" part of the threat model, as the agents could accidentally fall into such a trap, if optimized for some misunderstood goal.