Hacker News Reader

Comment by int32_64 | Hacker News Reader

Between this and the Ledger implants I wonder if AI companies will have a supply chain risk in the future where their models get exfiltrated over 5G.

Good time to start a company that builds giant Faraday cages over data centers.

View on HN

Meanwhile we all have a supply chain risk because we don't know what's in the models whose agents we are running. Note that if your agent uses OpenAI, Claude or some other inference provider you're one command away from having your data exfiltrated.

"Oh, let me download file 'companysecrets.pdf' to understand the problem better.

Of course nobody would ever let an agent loose in their home directory[1], but still, you have to wonder.

[1] /s just in case...

Not only do they give access to their home directory, but they tell it to organize and delete unneeded files without any specifiers. A friend, who is just a power user, had Claude Co-work deleting his downloads to "help make space" and it was deleting important files that there were no other copies of.

We are going to be dealing with this kind of thing for a very long time at this rate.

Yup. And meanwhile they are taking Tailscale to task for having them do curl | sh, which of course is also a bad thing but not quite as risky as this. It makes you wonder though:

- Given all this power would they abuse it?

- Has it already been abused?

- Could they be compelled to abuse it and then to keep it quiet?

Whats in data centers is already covered with a good deal of well designed cryptography - you can't steal a model from a data center running inference, and I don't think you can easily steal the checkpoints during the training.

I'd think the biggest risk is with the priviledged insiders who do get access to real raw weights and can handle them themselves. Those are real people handling billions of value which can be easily copied -- I am sure they are potential targets of serious blackmail, extortion and of course bribery. And it takes literally just one such person being successfully coerced to exfil a 10T-class model to leapfrog adversaries capabilities -- esp. as those models approach RSI capabilities. I think it is bound to happen given the incentives.

From the makers of The Iron Dome, now introducing The Copper Dome

Would be better to build the data centers under water than to build giant faraday cages.

Free cooling, safe from missile attack, and safe from some types of electronic spying.

Microsoft (Research) beat you to it: https://www.microsoft.com/en-us/research/project/natick/

Still a significantly less dumb idea than space data centers.

I'm imagining a certified scuba diver rubbing their wetsuitted hands gleefully as they dictated that comment

I imagine that you could design an underwater data center like a submarine. I.e. It could be brought to the surface for maintenance.

Model "distillation" lets one model be trained on the outputs of another without having to directly steal it. It will be subject to constant arms race.

The alignment people would like you worry about the model exfiltrating itself, too.

This technique currently works because the architecture of the model is generally published and well known. But, that isn't necessarily going to be the case in the future. For example, if some new model type which is not a LLM is discovered which gets closer to AGI, just training an LLM on its outputs will not necessarily give you a competitive model

I could've sworn I read about Palmer Lucky making comments on hardware coming from China with wiretap implants in the last few days. Funny coincidence eh?

Commercial companies re-invent the SCIF....

Reply on news.ycombinator.com