Skip to content
by InfinyAI ES

IA en un minutoNewsAgents

AgentsUpdated:

NVIDIA launches OpenShell to control what an AI agent can access

IA en un minuto newsroom · Editor: Jon Elgezabal

In 30 seconds

OpenShell is part of Open Agent Safety Platform, the open platform NVIDIA has built to protect AI agents from the moment they are tested until they are deployed. Its code is open and it sets which data and systems each agent may reach, and it is already available to everyone. Salesforce, SAP, Microsoft and Anthropic are among the companies backing the initiative, and Salesforce has already hooked it up to Slack.

Letting an AI agent into a company's data and tools now comes with an extra layer of control.

NVIDIA announced Open Agent Safety Platform on September 28. It is an open platform meant to protect agents all the way from testing through to deployment, and it includes OpenShell, an open-source environment that determines which systems and data an agent is allowed to reach, with the boundary placed outside both the model and the agent. It is already generally available and serves open and closed models alike: SpaceXAI uses the platform with Cursor's coding agents and with Grok models, and Anthropic has integrated it with Claude Managed Agents. Anthropic, Microsoft, Salesforce and SAP are among the companies joining the initiative.

NVIDIA's release adds that the platform comes with Sentry, a watchdog hosted on NVIDIA BlueField-4 DPUs that can quarantine, within milliseconds, an agent trying to step outside its limits. Working together, Salesforce and NVIDIA have built OpenShell into Slack, where teams can watch what each agent does and approve or refuse its requests for additional permissions. The software can be downloaded from NVIDIA's developer site and from GitHub.

Correction (6 October 2026): NVIDIA's release does not say OpenShell works with Codex and Claude Code or without rewriting the agent; the examples it gives are Cursor, Grok and Claude Managed Agents.

Why it matters · analysis and opinion

Putting the boundary outside the agent is what sets OpenShell apart: it is not the model that decides what it can reach, but a layer the agent does not control. If that boundary lived inside, an agent would only need to ignore its instructions to cross it. With Anthropic, Microsoft, Salesforce and SAP on board, and support for open and closed models, it has a chance to become a shared reference rather than one more item in NVIDIA's catalog. For an ordinary team, the most tangible part is Slack: seeing what each agent does and having a person approve or reject each request for more permissions. What remains unclear is how much it takes to run without NVIDIA hardware, since the Sentry watchdog runs on its DPUs.

Official source: NVIDIA · Written with the help of AI: how we make the news

AI in your inbox, every day or every Friday

The stories that matter, each one in a minute. With the source for every one.

Choose one or both:

Sign up and you are in: the daily arrives every night and the weekly on Friday mornings. You can unsubscribe from any email.