A sufficiently sophisticated AI could reorganize its thinking to appear innocuous while secretly pursuing a misaligned goal, making deception hard to detect