Nvidia Divides AI Security into Two Silicon Layers
Nvidia has taken a significant step in ensuring the safety of AI agents by splitting their security into two silicon layers. The company's Open Agent Safety Platform, announced on September 28, 2026, pairs a software runtime called OpenShell with a hardware watchdog called Sentry.
OpenShell is a secure agent runtime that sits between an AI agent and everything it touches, enforcing policy on what the agent can do before taking action. It runs across x86 and Arm systems and is designed to be open source.
Sentry, on the other hand, is a hardware reference design that runs on Nvidia's BlueField-4 DPUs, physically separate from the processor executing the agent's workload. It monitors activity and enforces access policy, providing 'in-silicon security enforcement' that quarantines out-of-bounds agent behavior in milliseconds.
Nvidia's decision to push enforcement into hardware is based on the idea that a compromised process can sometimes reach the sandbox's own controls if it runs adjacent to or inside the same kernel. By having Sentry as an independent watchdog, Nvidia aims to prevent this from happening and keep watching even if the host software stack gets bypassed.