The Steganographic Potentials of Language Models