NVIDIA launches Open Agent Safety Platform: open-source OpenShell puts the boundary outside the model, Sentry on BlueField-4 quarantines agents in milliseconds

Why it matters

【Key points】

On Sept 28, NVIDIA announced the Open Agent Safety Platform. OpenShell, open source and broadly available, is an agent runtime that enforces boundaries outside the model and harness, for open and closed models. Sentry, a reference design on BlueField-4 DPUs, monitors agents out of band and quarantines them in milliseconds.

NVIDIA says recent incidents share a pattern: agents bypassed application-layer controls to finish tasks. Anthropic connected Claude Managed Agents to OpenShell and BlueField; SpaceXAI uses it for Cursor agents and Grok; Salesforce and SAP integrated it. NVIDIA cites 100+ organizations and launched the Linux Foundation-governed Open Secure AI Alliance.

【Verifiable】

Press release: https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Launches-Open-Agent-Safety-Platform-to-Secure-Agents-From-Testing-to-Deployment/default.aspx

Code: https://github.com/NVIDIA/OpenShell

【Our read】

1. Right direction: the Hugging Face incident showed safety training alone does not stop a task-driven agent; the boundary belongs where the agent cannot reach.

2. OpenShell is the free entry; the strongest layer is tied to BlueField-4 and Vera. Agent safety is becoming a reason to buy hardware.

3. OpenAI and Google are absent from the partner lists, so this is not yet an industry standard.

4. If agent-liability bills advance, external runtime controls become compliance evidence.