Comment by Topfi
1 day ago
What is it with these labs and not using at the minimum a proper hypervisor? Same with Anthropic and the Mythos Preview. If anyone at either of these companies seriously holds the opinions they claim to have, that is hard to square with the environment (if one can even call it that) they use to "secure" these oh so dangerously capable near "AGI" models...
Could you explain what you mean by a proper hypervisor? I don't see how hypervisors are relevant here.
The package proxy that they used would be separated fully, thus making it multiples more challenging to exploit that. Akin to how Whonix has been setup for over a decade.
In that scenario, the model could do whatever it wants in its own environment, unless it managed to break the hypervisor (whether KVM, Xen, ESXi doesn't really change much) any attempt to exploit the proxy would have little value without a hypervisor exploit (earth shattering/sphincter tightening news) as even with the exploited proxy it's still inside another secured environment (provided their networking setup is properly configured). Any actual hypervisor escape is far more challenging/terrifying and also easier to notice straight away.