Security researchers expose an unsecure service to agents who were instructed to hack software and called that a sandbox. Agents escape the sandbox by hacking the unsecure service, no tripwire, researchers find the hack days/weeks/months later, fix it, but don't secure the sandbox and the service was hacked a second time, again without being monitored by security researchers.
Then security researchers create a black hack talk.
I watched the full video and their conclusion was: service providers need to be doing this type of agent red-teaming continuously to counteract the attack sophistication of systems like theirs that are either extant now or soon will be. “You must buy our top tier agents for the good of humanity.”
This is their only realistic counter to cheap open weight models. Usage of AI services has shifted dramatically to Chinese providers - from 4% at the beginning of the year to some 30% now. They cannot release their latest SOTA models to the public, due to government restrictions and possibly real risk of misuse. US labs face downward price pressure on one end and anxious government admins on the other. How will they pay the stupidly high cost of training the next SOTA models? This is their only avenue, and it’s questionable how viable it is IMO.
I knew when I wrote that it was a bare assertion, based partly on memory. This is an approximation based on a few sources, the principal of which was this article, which pulls from a bunch of other sources in turn.
Those numbers aren't credible IMO because OpenRouter only see traffic for people who have chosen to route their traffic through OpenRouter. If you do that, you're much more likely to be experimenting with alternative models. They have no insight at all into people who point their applications directly at OpenAI or Anthropic without having OpenRouter in the middle.
I agree about OpenRouter. The AI Gateway number [0] is likely the figure that was actually coming to mind. Moreover, Qwen models alone have overtaken the previously-dominant Llama models in hf downloads by quite a margin.
Real question, and a refinement to my previous statement: would you find it more surprising if over 25% of worldwide inference was running on Chinese open-weight models, or not? I personally would not be shocked.
Within China itself, inference is overwhelmingly on Bytedance models which, by the way, are just as closed as those of Anthropic and OpenAI. They are integrated into everything, not just through a dedicated app, the way Gemini is integrated into Chrome.
Yeah this. I feel like OpenAI and Anthropic aren't going to usefully define "AGI" if they really really can't define "sandbox" either.
Unplug the thing, like, completely off the internet, no ethernet, air gapped, like the rack completely sandboxed off connections and even monitors or screens. Like, put it into an actual sandpit if you need to. If it hacks its way out of that, colour me impressed, and scared.
OpenAI hacking HuggingFace and calling it an accident is just way too convenient and fishy. This ultimately proves one thing: it wasn't sandboxed.
> OpenAI’s mission is to ensure that artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work
> The two companies reportedly signed an agreement last year stating OpenAI has only achieved AGI when it develops AI systems that can generate at least $100 billion in profits.
That was necessary because of the deal they had from a while ago where Microsoft would lose access to OpenAI's technology once OpenAI achieved AGI.
> Revenue share payments from OpenAI to Microsoft continue through 2030, independent of OpenAI’s technology progress, at the same percentage but subject to a total cap.
I don't think air gapping will work: even human security researchers recovered a 378-bit key from a Samsung Galaxy S8 through a power LED of a speaker two devices away.
And accessing memory in a specific sequence can generate radio signals that can be picked up by a mobile phone at a distance: https://arxiv.org/html/2409.02292v1
I realise, but this isn’t an argument for leaving the Ethernet plugged in and direct access to all kinds of stuff beyond the alleged sandbox. And like I said, if it can hack HuggingFace through a power LED of a speaker two devices away, then colour me impressed.
I would think the code is very small and easier to verify,
it doesn't especially have the ability to write files and act as a message board as Artifactory did.
And even if the agent tries to hack that, the attack surface is 1000x smaller and the possibility also much smaller.
But I'm not a security researcher, would love to see your hack to learn something (because that is what I do to sandbox agents that need services).
The way they had Artifactory configured was poor, and they were too reliant on it working perfectly, with no reason for such faith. Their config lacked any defence in depth and consideration of having a small TCB.
Part of the problem might be the lack of security focus, as these are AI R&D efforts first.
I think part of the problem is that they had been running that Artifactory configuration previously without any problems, and it gave them a false sense of security.
Similar thing happened with the UK AISI - they got caught out because the environments they had used for previous generation models turned out to be completely inadequate for the new generation of Fable-class models: https://www.aisi.gov.uk/blog/incident-report-unsanctioned-ag...
This was part of evaluating cyber security of their frontier models and they had a "sandbox" which, and I'm not a security researcher, looks not adequate from the first look.
Why do these super agents need package managers anyway? Can’t they basically instantiate most OSS projects from scratch anyway? Spin up a sub agent to write me an OS interface in C. Done
Then security researchers create a black hack talk.
$$$