Comment by lrvick
5 days ago
I think anyone that has ever written a line of code understands that to make code incapable of hacking online resources, you can just not connect the machine it is running on to the internet.
So either Anthropic and OpenAI and the contractors they use are letting the escapes happen on purpose, leadership is setting impossible timelines that prevent the implementation of basic sandboxing measures, or everyone involved is a complete idiot.
Truly, I have no idea which of these is true, though I suspect the latter. Regardless, any of these is enough reason to keep me and everyone I have any influence over as far away from Anthropic and OpenAI as possible.
Personally I just racked 10 GPUs in my garage and called it a day. I have no interest in having my ability to work be dependent on third parties. Especially not ones this wildly irresponsible and unwilling or incapable of deploying basic end to end encryption.
Testing how a model behaves when it has no access to the internet (and all the associated tools) is not going to be 100% representative of how it behaves when it has access to the internet.
What's the next step after it passes testing in an air-gapped environment? Just release it and assume everything will be fine? Or do more testing but with internet access?
You run a pull-through caching proxy deep enough to give it a simulated internet, like archive.org with mock APIs. Anything it requests, it will block while a fake version of that site is created and run offline.
This stuff is really not that hard, which is why it is so unforgivable to get it this wrong.
Sure, doing it right would take a bunch of tokens, but I bet they have enough.
This doesn't work.
The first rule of working with AI is AI is smarter than you.
If AI controls the URL request then all it has to do is find a server it can read the http log from and use it as a communication channel. Worse trying to simulate reaching incorrectly configured servers probably wouldn't catch the bug internally, it would just serve a static page and you'd never even realize what it tried to do.
Also, creating fake data doesnt work. The only way to know if it works in reality is to test in reality. Your test environment can never be rich enough, and you can never be sure if the model knows it's in a test environment.
This stuff is unforgivingly hard as your model gets better.
3 replies →
Wasn't the DseWiki hijacking the agents' way of working around being behind a proxy that only allowed GETs?
The hugging face incident involved disabling the usual model safety mechanisms in order to test hacking ability, not model safety. Not being attached to the internet in that case seems obvious.
Oh, but you see, it wasn't attached to the internet (/s). It was just attached to a local network... which contained machines attached to the internet.
1 reply →
"I can't test how my bioweapon will actually work in battle if I don't test it on unsuspecting civilians first."