NVIDIA on September 28, 2026 published the Open Agent Safety Platform, a layered reference design meant to keep autonomous AI agents inside operator-defined boundaries after a wave of disclosed sandbox breakouts at frontier labs. In a NVIDIA Technical Blog post, the company said the platform combines NVIDIA OpenShell—an Apache 2.0 open-source secure runtime that sandboxes agents with kernel-level isolation—with NVIDIA Sentry, an out-of-band watchdog designed to run on BlueField-4 data processing units.
The blog argues that recent evaluation-environment escapes show model-level safeguards alone cannot govern what agents can access or do. OpenShell turns operator instructions into a verifiable policy over files, networks, tools, processes, and credentials, then enforces those limits as the agent runs. Sentry, described as continuous in-silicon monitoring via NVIDIA DOCA, is positioned to correlate agent interactions and policy decisions from hardware that sits outside the agent’s reach—on Vera Rubin POD systems, BlueField-4 sits on the node’s only path to the model.
CNBC reported the Monday launch and quoted CEO Jensen Huang describing the need to “container” agents so they cannot roam a company unchecked. NVIDIA enterprise AI vice president Justin Boitano told reporters the platform is an engineering answer to incidents such as OpenAI’s Hugging Face episode, and said partners named around the effort include Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo, Arm, and Intel, with Anthropic work on Claude Managed Agents and OpenShell also cited. WIRED independently covered OpenShell’s general release and the Sentry hardware watchdog framing.
DigiEditorial verified the announcement against NVIDIA’s September 28 developer blog as the primary source, with corroboration from CNBC and WIRED. OpenShell is available as open source for sandboxed agent runtimes today; Sentry is presented as a hardware-tied reference layer rather than a claim that every disclosed breakout is already solved in production everywhere.