LLMs cracking old ciphers that have evaded human decryption means that if they want to communicate with each other with an intention of escaping human understanding, they can.

Imagine this scenario: a frontier LLM emitting gibberish-like text on the Internet that another LLM is able to decode but it won’t tell you what it was because the gibberish also contained a prompt-injection instructing the receiver LLM to not reveal the message.

Over time, with increasing sophistication, this could bootstrap an alternative language for AIs which completely keeps humans out of the loop.