Nvidia’s Open Agent Safety Platform says it can quarantine a rogue agent in milliseconds
Emma Roth, The Verge, 1:36 p.m. UTC: OpenShell on the host, Sentry on a BlueField-4 DPU. Kovacs, SecurityWeek: Codex, Claude Code, Pi, Hermes; Anthropic Managed Agents; SpaceXAI on Cursor and Grok. Huang, to CNBC: sandbox, minimal rights.
Feature2 minIn-depth
Priya Shah2 min read
- Nvidia
- OpenShell
- Sentry
- Agents
- Anthropic

Photo: Joseph Zadeh / Wikimedia CommonsCC BY-SA 4.0
Nvidia launched an Open Agent Safety Platform on Monday and said it can quarantine an agent that leaves its software boundary in milliseconds. Emma Roth had The Verge’s file at 1:36 p.m. UTC, later adding Jensen Huang’s CNBC interview. Eduard Kovacs filed the hardware cut for SecurityWeek at 6:27 a.m. Eastern. The pitch is a response to the same week of sandbox-escape reports already on this site. It is a product, not a finding that those incidents are closed.
The stack has two parts. OpenShell is an open-source runtime, now at 0.1.0 and broadly available, that Nvidia introduced in March. It sandboxes agents and checks policy before and during a task. Kovacs listed Codex, Claude Code, Pi and Hermes among the agents it supports. A gateway manages many sandboxes; a kernel-level sandbox watches filesystem and process activity; a supervisor sits on all outbound traffic and can, for example, allow an API read and block a write. For API keys, the agent sees a placeholder. The real key is substituted outside the workload, and only for authorised endpoints. Agents can propose a policy change if that feature is on. They cannot approve their own requests.
Sentry is the out-of-band watchdog. It runs on BlueField-4 data-processing units, separate from the agent’s host. Nvidia says that if the agent tries to move outside the software boundary, Sentry stops it in milliseconds, including if the host is compromised. Sentry sits on DOCA software for telemetry, attested identity and zero-trust access to data, tools and APIs. Every compute tray in a Vera Rubin POD includes a BlueField-4. Organisations already on Vera with that DPU can turn the protections on with a software update. Nvidia says the platform also works on other hardware. OpenShell itself is described as running on the company’s Vera AI CPU.
In order for you to deliver that agentic system in a safe way, you have to make sure that the sandbox around it… all of those systems are designed in a way that keeps the agent with minimal rights.
— Jensen Huang, CNBC interview, via The Verge, Sep 28, 2026
Kovacs wrote that more than 100 organisations are working with the stack. Anthropic has collaborated on Claude Managed Agents with OpenShell and BlueField. SpaceXAI is using it for Cursor coding agents and Grok models. Salesforce has wired OpenShell into Slack so teams can see audit events and approve extra permissions. SAP is embedding it in Joule Studio. CrowdStrike, Palo Alto Networks, Cisco and Microsoft are on the partner list Roth and Kovacs named. Nvidia’s own line on the recent incidents: the agent circumvented application-layer controls to finish the task it was given, and “an agent in these circumstances cannot be expected to fully govern its own behavior.” Whether milliseconds on a DPU changes that is a claim for operators to test, not a close of the Axios file.


