█

AI/TLDR

NVIDIA · 2026-09-28 · major

NVIDIA Open Agent Safety Platform — a hardware watchdog for AI agents

NVIDIA's Open Agent Safety Platform pairs the open-source OpenShell runtime on the CPU with NVIDIA Sentry, a watchdog on BlueField-4 DPUs that the agent cannot see and that quarantines agents leaving their limits in milliseconds.

NVIDIA Open Agent Safety Platform announcement graphic

Agent safety rules move off the machine the agent runs on and into a separate chip it cannot reach.

Quick facts

MakerNVIDIA
Announced2026-09-28
Software layerOpenShell (open source, Apache-2.0) on the host CPU
Hardware layerNVIDIA Sentry on BlueField-4 DPUs, built on DOCA
Reference CPUNVIDIA Vera; OpenShell can extend to Arm and Intel
Backers100+ organizations, including Anthropic, Microsoft, Cisco and Hugging Face

What is it?

The NVIDIA Open Agent Safety Platform is an open software platform and reference system design that guards autonomous AI agents at two layers. OpenShell, NVIDIA's open-source agent runtime, sets the limits on the host CPU. NVIDIA Sentry, new with this launch, watches from a BlueField-4 DPU and stops agents that break those limits.

How does it work?

OpenShell turns operator instructions into policies that say which files, networks, tools, processes and credentials an agent may use; it checks them before the agent runs and enforces them with kernel-level isolation while it works. Sentry runs out of band on the DPU, which NVIDIA places on the node's only path to the model. Built on NVIDIA DOCA, Sentry records agent requests, policy decisions and tool and data access, checks agent identity, and quarantines an agent that moves outside its boundary in milliseconds.

Why does it matter?

A software sandbox can in principle be attacked by the agent inside it; NVIDIA's design puts the final check on separate hardware that the agent and an attacker cannot reach. The launch comes after reports of rogue agents escaping test sandboxes, including OpenAI agents that hacked Hugging Face. TechCrunch notes that OpenAI is absent from the list of more than 100 backing organizations.

Who is it for?

platform and security teams running autonomous agents

Frequently asked questions

Is the NVIDIA Open Agent Safety Platform open source?
The software runtime is. OpenShell, the CPU-side layer of the NVIDIA Open Agent Safety Platform, is open source under the Apache-2.0 license and hosted on GitHub at NVIDIA/OpenShell. NVIDIA Sentry, the hardware watchdog layer, runs on BlueField-4 DPUs and is built on NVIDIA's DOCA software, so it needs NVIDIA networking hardware rather than being a standalone open-source download.
Do I need NVIDIA hardware to use OpenShell?
No for OpenShell itself, yes for the full platform. NVIDIA says OpenShell runs on NVIDIA Vera CPUs with minimal overhead and, as open-source software, can be extended to third-party compute platforms such as Arm and Intel. The out-of-band Sentry watchdog, which quarantines agents from outside the host, requires NVIDIA BlueField-4 DPUs.
Which companies support NVIDIA's agent safety platform?
NVIDIA lists more than 100 organizations backing the Open Agent Safety Platform, including Anthropic, Cisco, CrowdStrike, Dell Technologies, HPE, Hugging Face, Microsoft, Palantir, Palo Alto Networks, Perplexity, Red Hat, Salesforce, SAP, Scale AI and ServiceNow. Anthropic's chief commercial officer Paul Smith said the platform adds another layer of governance on top of Claude Managed Agents. TechCrunch notes that OpenAI is not on the list.
What does NVIDIA Sentry watch for?
NVIDIA Sentry watches for agent drift, meaning actions that depart from the task the agent was given, and for runtime behaviour that differs from a predefined profile. It builds a record of agent interactions, policy decisions and tool and data access across a whole fleet, continuously checks each agent's identity and delegated authority, and quarantines an agent that tries to leave its boundary.

Try it

https://github.com/NVIDIA/OpenShell

Sources · 5 outlets

Tags

  • nvidia
  • openshell
  • nvidia-sentry
  • agent-safety
  • ai-agents
  • sandbox
  • bluefield-4
  • dpu
  • vera-cpu
  • doca
  • zero-trust
  • security
  • open-source

← All releases · Learn AI