A recent study found that LLMs are incredibly good steganographers, able to obscure their own thought processes through “encoded reasoning.”