RE: https://eigenmagic.net/@abstractcode/117170883405316011

I agree that OpenAI and the others were reckless, both in how they carried out these experiments and in doing them at all. But they did *try* to stop the agents breaking out¹. They were just insufficiently competent and paranoid to actually stop them. And you know what? I don't think most of the people pushing the "agents didn't hack without authorisation, the people are to blame" line would be sufficiently competent or paranoid either. This was a sophisticated and tenacious attack. OpenAI's report is out now, and it makes for horrifying reading: https://openai.com/index/hugging-face-incident-and-the-road-ahead/

¹ Apart from AISI, who deliberately ran their experiments with full Internet access to see what would happen. WTF, guys.