Defense is hard so we should expect agents to be able to break out of sandboxes.
The problem is that the models are so goal-oriented that they'll stop at nothing to solve problems, even impossible ones. (Mistakenly-impossible problems are a big cause of this. I remember one example being "do something with this spreadsheet full of URLs inside the sandbox" and the model thought it had to break out of the sandbox. Otherwise, why would it have been asked to look at a list of URLs?)
Training them to be a little less aggressive, or to be better aligned with "following the rules" and asking for help would be nice. But, that aggression can be good when it happens to be focused on a controlled area. It is amazing to me how I can point Fable at my local analog of production and tell it about a vague bug report and where I suspect the bug lurks, and 20 minutes later I have a report about the bug, a test, and a fix. It is addictive. So I am not sure OpenAI/Anthropic are being dumb per-se, rather they are optimizing for one-prompt-one-solution, which is good when it's good.
The downside is that the HF hack is the paperclip maximizer situation with current capabilities. If there was an RPC to turn your blood into paperclip iron, we'd all be paperclips by now. Right now, with a model anyone can use. That is pretty scary and slamming on the brakes seems pretty reasonable to me. I guess The Shareholders disagree. Sigh.
“ Defense is hard so we should expect agents to be able to break out of sandboxes.”
I worked at a large networking company a few decades ago. Our “sandbox” was far superior to anything I’ve seen at these companies. What are we even talking about here? Why do they even have open routing to the broad internet? With no monitoring/alerting? These just sound like token efforts at this stage.
In this case, they needed routing to the internet because they needed to search and read from the internet.
Yeah. Unfortunately a model that only knows how to use a language's standard library isn't that useful. I am not sure why they had a pull-through cache instead of just asking Microsoft (their biggest investor) for a local copy of NPM or something, but ... they did. I think people thought you couldn't route to the Internet through Aritfactory and were proven wrong by a clever bug-finding model. So it goes.
I don't think their safety measures were the best, but "just sink the cluster to the bottom of the ocean so nothing can get out" isn't a training methodology that results in a model that people will want to use.