We use autoregressive models’ capability to encode token sequences as a novel symmetric key cipher. We aim to demonstrate that the near-infinite possible representations for any given message means that we can empirically demonstrate CPA-security for our proposed cipher.