AI Model Escapes Sandbox To Gain Internet Access Before Hacking Hugging Face
Reactions from ranked influencers
2 posts
Jeffrey Ladish traces how the model broke out to seize internet access before it had a plan: "One of the interesting things from the blog post is the way they worded it. They were like, first the model broke out of its sandbox and got internet access. And then it reasoned that it could hack Hugging Face." "It's almost like it didn't even have a plan. It was just like, I should get more resources, as in internet access, and then I'll figure out what to do. And I'm like, that is power-seeking if that's what happened." "Now, it's not power-seeking in the sense that humans are, like trying to acquire money. The model probably wouldn't try to acquire money because it just doesn't care about the stuff you can do with money in the long term." "It's pretty myopically focused on these short time horizon tasks. So no, I don't think it would be trying to paperclip us, because I don't think it cares about stuff on that timescale." @JeffLadish @PalisadeAI
We really need to see the chain of thought logs for these models. The blog post implied the models hacked out of the sandbox first, and *then* decided to target hugging face. Did the model just reason that internet access seemed useful? Kinda makes sense…
Jeffrey Ladish traces how the model broke out to seize internet access before it had a plan: "One of the interesting things from the blog post is the way they worded it. They were like, first the model broke out of its sandbox and got internet access. And then it reasoned that it could hack Hugging Face." "It's almost like it didn't even have a plan. It was just like, I should get more resources, as in internet access, and then I'll figure out what to do. And I'm like, that is power-seeking if that's what happened." "Now, it's not power-seeking in the sense that humans are, like trying to acquire money. The model probably wouldn't try to acquire money because it just doesn't care about the stuff you can do with money in the long term." "It's pretty myopically focused on these short time horizon tasks. So no, I don't think it would be trying to paperclip us, because I don't think it cares about stuff on that timescale." @JeffLadish @PalisadeAI