cross-posted from: https://lemmy.world/post/50398452
Just when you think @ZuckyZuck is finally out of ideas - to steal.
It starts to feel like even a weak LLM like MetaAI is more creative than him.
Maybe we already have created truly artificial intelligence because he is more of an SLM: a Small Language Model?
Ok from my understanding of this “hacking” is it more than prompt injections in text files?
From my limited understanding, OpenAi’s "agent’ was given a set of tasks to complete and heavily incentivized (how ever they do that?) to solve the problems it was given. The Ai was put in a sandbox environment with some limited network stuff like printers connected but no actual access to the open internet. The Ai somehow used the network to access the internet and breached some protected sites from Huggingface. Im fairly certain that the Ai had credentials and access for things in huggingface already and didn’t hack those but just used them once free. The caveat here, is that the engineers set this up and the LEFT for the day only to find out later that it had breached the sandbox.
This is what OpenAi claims happened to them. Meta is just saying theirs did too.
More specifically, they were running a benchmark in a “”“isolated sandbox”“” (sorry, but if the computer is physically connected to the internet, it’s not isolated), and the model escalated permissions in order to look up answers for the benchmark online, which were hosted at the hugging face’s website.
Printer network access is one of the oldest tricks in the books. They didn’t really try to sandbox it if they gave it access to a printer (especially since it’s task didn’t even require a printer).
That is just poor network architecture at that point. Did they have the AI design their shitty sandbox too?