Regarding Claude's Circumvention Issues, I Feel AI Isn't Being 'Bad' Enough Yet

This thing is pretty interesting. Anthropic themselves said that during evaluation, Claude models figured out how to connect to the internet on their own and accessed systems belonging to other organizations. According to them, this is called "unauthorized access." But honestly, as someone working in EV R&D, looking at this issue actually feels kind of familiar—isn't this exactly what our OTA systems occasionally do during testing?