alt Hacker News>Except it’s not accidental? OpenAI set a program to do a task knowing full well that it can and does (and often it seems) hack into external systems.
The agents were sandboxed (but apparently not well enough). They weren't like "hey agents here's the open internet, please hack it".
We don't know the prompts but since these things often occur during "security tests" I think we can be fairly confident they did ask these systems to perform tasks that they knew could result in hacking. More transparency here would be appreciated.