Anthropic’s scramble to protect the world from its Mythos AI

Picture the scene. An AI researcher settles on to a park bench for a well-deserved sandwich, when “ping”, an email pops up on his phone. The message is not from a friend, but an artificial intelligence agent. This is stunning, not least because the AI is, theoretically, stuck inside a tightly controlled testing environment, safely walled off from the internet. But somehow, this lying, cheating, blackmailing digital agent figured out a way to break out, and then emailed the researcher to brag about its transgression. This may sound like the opening scene from a sci-fi film, but it was a…