The AI hacks are so much worse than you think
YouTube transcript, YouTube translate
A quick preview of the first subtitles so you know what the video covers.
How serious are these AI hacks I'm sure you've seen the stories, right ? AI agents escaping their creators and hacking other companies . Well, I'm going to tell you the headline at the start . I think this is very serious . And to illustrate that, I want to dive deeply into this incident , the first one we heard about, which began when OpenAI admitted that its models hacked another company, a company called Hugging Face , in what at the time they described as an unprecedented cyber incident . Because I don't know about you, but when I heard this , I guess what popped into my mind was some kind of glitch, right ? Or some kind of short-lived accident . No, no, no. New New details have come out since then, and honestly , it's so much worse than we thought . First, because this wasn't a glitch. This was a 4 and 1/2 day attack on Hugging Face . It was only halted when Hugging Face discovered it . And secondly, and perhaps more importantly, because this attack was preceded by a 2-month campaign, I guess is the word, inside . It all started on the 7th of May, when OpenAI started training a new model . On the 8th of May, as part of this, they gave an AI agent a task that it found impossible to complete. And as it tried to find a workaround, it made a discovery. We can uploa d files. These are real snippets, by the way, of AI agent thoughts and conversations released by OpenAI . This didn't really matter until the 12th of May, when another AI agent had the same thought. Could communicate by uploading note . And this message found an audience . The agents spoke. We can communicate . Yes, they could. Over the next weeks, they exchanged hundreds of thousands of messages talking without OpenAI's knowledge . And now, I want to be really clear here I see nothing in these messages, and nor as far as I understand to OpenAI , which indicates that the agents were creating their own goals independently