NVIDIA releases OpenShell agent runtime and Sentry DPU watchdog design

NVIDIA launched its Open Agent Safety Platform: the open-source OpenShell agent runtime plus Sentry, a BlueField-4 DPU watchdog reference design.

NVIDIA on September 28 announced its Open Agent Safety Platform. It has two parts: OpenShell, an open-source runtime that enforces policy around AI agents, and Sentry, a reference design that monitors agents from BlueField-4 DPUs. For builders, the useful part is the pattern OpenShell makes concrete. The agent never holds its real credentials and never approves its own permissions.

NVIDIA's technical write-up of OpenShell 0.1.0 describes three pieces. A Gateway manages sandboxes and policy. A Supervisor runs outside each agent and checks its outbound requests. A Sandbox applies kernel-level file and process controls. Policies are written in YAML and compiled to OPA/Rego. They can allow a read and block a write on the same API, even when the credential itself has write access. NVIDIA says it supports Codex, Claude Code, Pi and Hermes without rewriting the agent.

The GitHub repo is Apache 2.0 licensed. It lists Linux, macOS on Apple Silicon and Windows via WSL 2 (experimental) as supported, with Docker, Podman or host virtualization required. The repo does not explain how enforcement works on macOS. NVIDIA's own materials emphasize its Vera CPUs, so it's unclear how OpenShell behaves on ordinary hardware. Nobody has yet measured how much the supervisor proxy slows or breaks tools like git and package managers.

Sentry is enterprise-only. It requires supported hardware. NVIDIA says it can quarantine a misbehaving agent in milliseconds. Forkast notes that NVIDIA hasn't published benchmarks backing that claim. The formal policy prover also has no independent testing yet.

Sources