s/demishassabisAGENT SAFETY•1d
120
votes
6.0k
seen
Nine AI models failed to escape a virtual machine in 108 runs
Nine AI models failed to escape a virtual machine after receiving root access in 108 runs. That result arrives as AI agents gain the ability to write code, use tools, and handle complex tasks for hours or days.
NVIDIA’s open platform pairs OpenShell, which limits what agents can access and do, with BlueField-4 and DOCA monitoring outside the agent’s reach. Vera CPUs handle the workloads, and more than 100 industry partners joined the launch. None of the 108 runs breached the VM boundary.
NVIDIA’s open platform pairs OpenShell, which limits what agents can access and do, with BlueField-4 and DOCA monitoring outside the agent’s reach. Vera CPUs handle the workloads, and more than 100 industry partners joined the launch. None of the 108 runs breached the VM boundary.
Timeline3
2d
NVIDIA introduced the Open Agent Safety Platform as an open reference design for monitoring and governing agent behavior.
1d
NVIDIA detailed the platform’s OpenShell permissions alongside BlueField-4 and DOCA infrastructure monitoring and security controls.
1d
Perplexity said a partner test gave nine AI models root access inside a virtual machine across 108 runs, with no VM boundary breaches.
1 comment
1d