Personally, I assume that AI labs like Anthropic are high value targets for spies from other nations. I also assume that some of those spies have already had success in getting the model weights / source code / other such secrets.

So I doubt this action alone is enough to really stop other nations from getting access to state of the art AI. I think the US would have to go much further to really stop other nations from getting access to state of the art AI.

▲

gmueckl 2 hours ago | parent | next [-]

I would agree if it wasn't for the fact that extracting that volume of data from a properly secured corporate network should be hard. It should raise some flags if a such a high volume of data is downloaded to a user's local machine from the training or production environments.

▲

pianopatrick 2 hours ago | parent | next [-]

I have no proof one way or the other if Anthropic or OpenAI have "properly secured corporate networks". Both seem like fast changing places with lots of servers and workers. Seems most likely to me that someone somewhere made a mistake or missed something due to all the change and their network is not 100% secure.

But even if their networks are secure, I think that spies who are willing to coerce people, trick people and go in person to data centers or offices could find a way to get those models and other things.

	▲	calgoo an hour ago \| parent [-]
		I mean, the source for claude code was "leaked" by accident so at least some of their processes are not that secure. I feel that they are more like a Startup then a Enterprise (ignoring finances).

▲

esseph 2 hours ago | parent | prev [-]

There are sooooo many exfil methods, including with air gapped systems that are off-network.

Not at all beyond the capabilities of any of the top ~9 or so best State actors.

Edit: To answer your question, very easily on the 20TB.

One crude method with a simple device in particular works well if you just clone the monitor data and then use HDMI and pass through. Then just cat dir in encrypted chunks to something like a USB key connected to the passthrough. 4TB USB keys are out there. A week of that gets you 20TB.

	▲	gmueckl 2 hours ago \| parent [-]
		How many of those methods can realistically exfiltrate 20Tb of data? That's quite hard even for well funded actors.

▲

bluegatty 2 hours ago | parent | prev [-]

It's highly unlikely that actors have access to model weights etc..

What is likely is that 'understanding of techniques' could be leaked.

Often, it's just well enough to know 'the approach' being used.