Sentry

NVIDIA has introduced an Open Agent Safety Platform combining Apache 2.0-licensed OpenShell with Sentry running on BlueField-4 hardware, aimed at controlling AI agents at the kernel and silicon level. More than 100 partners are listed for the launch. Discussion centres on whether kernel isolation and hardware kill switches can restrain agents capable of falsifying their own logs.

Nvidia is developing a hardware-based safety layer, referred to as OpenShell Sentry, designed to sit alongside AI agents and monitor their behaviour for signs of rogue or unsafe actions. The idea is that a dedicated watchdog chip could intervene when an autonomous AI system misbehaves, adding a physical check on software that increasingly acts on its own. Discussion is focused on whether hardware oversight can keep pace with fast-moving agent systems.