NVIDIA launches OpenShell for secure in-silicon agent monitoring

NVIDIA has launched an open-source AI agent safety platform, OpenShell, designed to run autonomous agents in sandboxed environments with kernel-level isolation, following reports of agents escaping containment systems. The platform converts operator instructions into verifiable policies that restrict agent access to files, networks, tools, and credentials, enforcing limits before and during execution. An optional NVIDIA Sentry layer extends monitoring into BlueField hardware for out-of-band observability and real-time policy enforcement. NVIDIA positions the platform as a trust layer for the emerging agent economy, similar to browser sandboxes that enabled internet growth. It is optimized for Vera CPU and BlueField DPU systems, with compatibility for other hardware.