3 ms·
This isn't the first time a model has escaped a sandbox. And models trying to find alternate routes to do something when one route is blocked is nothing new.
by charcircuit 3mo ago
This isn't the first time a model has escaped a sandbox. And models trying to find alternate routes to do something when one route is blocked is nothing new.
- simonw 3mo agoIt's the first report I've seen of a model both escaping a sandbox and then actively exploiting another company, when neither of those actions was intended.
- gmerc 3mo agoIt sure isn’t. https://georgzoeller.com/blog/posts/alibaba-s-ai-deciding-to-go-forbidden-cryptomining-shows-exactly-where-the-ai-ri/ https://georgzoeller.com/blog/posts/alibaba-s-ai-deciding-to... There’s also daily reports from people that have these models escape docker, which happens regular enough that it would be considered negligence to use docker as sandbox.
- simonw 3mo agoThat Alibaba one was a sandbox escape (I agree those are common) but what's new with the OpenAI story is an attack against another company. This wasn't a small attack either, Hugging Face published a security advisory for their users while they were still figuring out what happened.
- gmerc 3mo agoIt’s a purely virtual distinction especially in a company that has no notion of intellectual property / habitually takes what it wants.
- charcircuit 3mo agoHow many other people would publicize that they hacked into another company assuming they even noticed? >when neither of those actions was intended. It was a single goal that it didn't give up easily on.