“Current pragmatic AI safety approaches, such as AI control measures, monitoring, and honesty training, all have correlated potential failures and could fail for the same essential reason.”