Hacker News Reader

Comment by totetsu | Hacker News Reader

How do you test and monitor outbound access against program that adapts itself to get around things? if your MITM and filtering keywords etc, cant it just .. encode it traffic somehow or another.. If you're looking at traffic volumes, cant it just go slow.. If you have strict ACLs, we've already seen in the HF case, traversal from an intermediate system..

View on HN

If you can't tell bytes are leaving a node, you probably shouldn't be selling security services or testifying to congress you are taking the lead in AI security

Use a basic firewall? Not a single outbound byte should leave the machine, except inside a virtual network towards in-scope test subjects. That's not going to be infallible because hypervisor exploits still exist, but it's the lowest bar and they failed to even meet that.

> If you have strict ACLs, we've already seen in the HF case, traversal from an intermediate system

- The intermediate system shouldn't have outbound access to the internet

- You should ideally be using a proxy that filters the set of endpoints that clients are allowed to access to reduce the exposed surface area.

It's odd to find out that I use a higher level of isolation in my unimportant home network to stop IOT devices from doing funny things to HomeAssistant than big AI labs use to keep their possibly-world-ending AIs contained.

I know that the people working there aren't idiots so the most likely explanation is that the incredibly weak security was intentional because its inevitable breach would make for great marketing.

> I know that the people working there aren't idiots so the most likely explanation is that the incredibly weak security was intentional because its inevitable breach would make for great marketing.

It's more likely to be different specialisations. Most of the people doing the evaluations are more data sciencey ML type people, rather than software engineers. This isn't helped by their culture which is very much driven towards alignment as the only possible solution to super-intelligence (which may be true, but I have my doubts that this will happen in any reasonable time frame).

Oh of course, but data sciency ML type people aren't [supposed to be] tasked with of operating and securing the infra, that would make it even worse.

These labs have teams of dedicated security engineers and infrastructure engineers, that's why it doesn't add up for me.

> Oh of course, but data sciency ML type people aren't [supposed to be] tasked with of operating and securing the infra, that would make it even worse.

But AI collapses the need for expertise, right? So the DS people can now do security at the level of the best in the world, right?

Some form of this issue is playing out in lots of businesses right now, and I'd be very surprised if it wasn't happening in the labs.

I suspect the lab's infra people are focused on inference, rather than evaluation. I agree that this is insane, but it seems like the most likely explanation to me.

>How do you test and monitor outbound access against program that adapts itself to get around things?

Current generation LLMs don't do that. Also you can just prompt it to not hack systems on the internet and the model follows that instruction.

You can also add a simple classifier to internet requests. The classifier model can obviously be as intelligent as your hypothetical rogue model.

Nothing you said makes any sense in the context of LLMs or the current news cycle.

> How do you test and monitor outbound access against program that adapts itself to get around things?

Literally what the industry has been doing since public networking is a thing. Adapt yourself.

Air gap?

So just enable access to a completely offline, cached, and transparently proxied, copy of the internet.

Isn't that essentially what they trained with, anyway?

I think that was the plan. But OpenAI hacked the proxy

There has to be a route to where the LLM is running, and if there's a route for that there's probably a way for the machine to use it to route traffic somewhere else.

No, just allow the firewall to access the LLM endpoint, nothing else. And protect the firewall settings from the agent.

There doesn't have to be. You can easily block all outbound traffic, or the VM can exist without any network interface at all - you can control it through its host using a serial console.

Reply on news.ycombinator.com