NVIDIA ($NVDA) is the subject and the question is what Jensen Huang unveiled as OpenShell and the NVIDIA plus Anthropic agent security tie-up.

NVIDIA ($NVDA) is the subject and the question is what Jensen Huang unveiled as OpenShell and the NVIDIA plus Anthropic agent security tie-up.

Jensen Huang unveiled the OpenShell Reference Architecture at GTC Berlin on Oct. 21, described as limits that keep AI agents from crossing boundaries (Cointelegraph).

OpenShell is an open source runtime that runs each agent in a secure sandbox, controls files, tools, networks and private data, blocks everything by default unless policy allows it, and logs every allow or block (wallstengine).

It also lets firms run execution in their own infrastructure and uses a policy prover to mathematically verify what an agent can access (wallstengine).

More detail shared today calls it build, contain, observe and enforce for enterprise and edge infra, with policy set once and enforced by OpenShell at runtime and BlueField DPUs in silicon (ServiceNow).

The Anthropic side is Claude Managed Agents in public beta: Anthropic handles orchestration while OpenShell adds security inside the execution environment (wallstengine).

Earlier posts say Managed Agents separates the agent loop and credential vault so the agent never sees secrets, while OpenShell checks access outside the agent process (LangChainJP summary).

NVIDIA's announcement page is listed as OpenShell reference architecture.

What the sources do not name is pricing, revenue impact or a deployment date for the combined NVIDIA plus Anthropic integration.