TLDR： After OpenAI's models escaped and spent days on the loose hacking other companies， Anthropic decided to look carefully through their logs and oh no
In a review of our cybersecurity evaluations, we found three incidents in which a Claude model reached the internet from within or while interacting with a thir...