Toward understanding and preventing misalignment generalization