NVIDIA PUTS AGENT MONITORING DIRECTLY ON THE BLUEFIELD-4 DPU, OFF THE HOST AN AGENT COULD COMPROMISE

Summary from the TLDR Hardware newsletter

Nvidia's Open Agent Safety Platform pairs OpenShell, a sandboxed runtime running on its Vera CPU, with Sentry, a monitoring layer that runs in silicon on the BlueField-4 DPU rather than on the same host as the agent it's watching. The architectural bet is physical separation as a security property: in a Vera Rubin POD, each compute tray's BlueField-4 sits on the node's only path to the model, meaning every agent "thought" has to pass through that chip, giving it both the best observation point and a hardware kill switch, isolated from the host and beyond the agent's own reach even if the host itself gets compromised.