NVIDIA launched Open Agent Safety Platform, combining OpenShell runtime controls with Sentry monitoring on BlueField-4 data processing units.

NVIDIA announced Open Agent Safety Platform as an open software platform and reference-system design for governing AI agents from testing through deployment. Its two principal layers are OpenShell runtime controls and Sentry out-of-band monitoring, rather than a new AI model or a GPU-compute service.

OpenShell establishes enforceable boundaries for agents running on CPUs and is reported broadly available. The announcement says it runs with very low overhead on NVIDIA’s Vera CPU and, as open-source software, can extend to third-party Arm and Intel computing platforms; these are platform claims without quantified overhead measurements in the supplied excerpts.

Sentry runs on BlueField-4 data processing units (DPUs), continuously monitors agent behavior and independently enforces safety policies in silicon. NVIDIA says it can isolate and stop an agent within milliseconds if the agent attempts to cross software boundaries; this is an announced capability, not a separately supplied benchmark.

Anthropic’s collaboration adds Claude Managed Agents, which separates the agent loop from execution sandboxes on different servers and centrally holds credentials in a credential vault. Reports say OpenShell blocks agent actions by default and identify Notion, Rakuten and Asana as production users of Claude Managed Agents; that adoption does not by itself establish production deployment of the entire BlueField-4/Sentry reference system at those companies.