NVIDIA's Agent Safety Watchdog Runs Outside the Agent, Which Is the Right Place

On 28 September NVIDIA announced the Open Agent Safety Platform, an open platform and reference design for securing AI agents from testing through deployment. It has two parts. OpenShell is open-source runtime software that enforces policy and traces agent actions on NVIDIA Vera CPUs. Sentry is a reference design that runs on BlueField-4 DPUs as an out-of-band watchdog, and NVIDIA says it can quarantine an agent that tries to leave its boundaries in milliseconds.

Our take: out-of-band is the key phrase. A safety check that lives inside the same process as the agent can be talked around, or fail along with it. A watchdog on separate hardware cannot be persuaded by a clever prompt, because it never reads the prompt.

NVIDIA says more than 100 organisations are working with the platform, including Anthropic, Microsoft, CrowdStrike and Cisco. OpenShell is broadly available now, while the release gives no separate availability date for Sentry.

The lesson travels well beyond data centres. When you give an agent access to a client's site, files or accounts, put the limits somewhere the agent cannot edit. Permissions, scoped tokens and audit logs do that job on a small scale.