This is like how the Stuxnet worm “escaped containment” back in 2010.
OMG! This proves that Stuxnet was conscious! The bugs were signs of super-intelligence! Everybody throw billions of dollars at Stuxnet worm to usher in a new Renaissance!
SO STUPID.
Really important to note that in no way shape or form did Anthropic’s models in this incident “escape containment,” and not just in the word-policing sort of way. The model stumbled through an open, misconfigured gap. From Anthropic’s incident report: