PRISON: Unmasking the Criminal Potential of Large Language Models