AI in a vat: Fundamental limits of efficient world modelling for agent sandboxing and interpretability