NVIDIA Open Agent Safety Platform

NVIDIA's open hardware-and-software stack for containing AI agents, pairing a CPU-level runtime with a DPU watchdog that can quarantine rogue agents in milliseconds.

NVIDIA · Security and safety tooling · security

Published in Feed

Overview

NVIDIA announced the Open Agent Safety Platform on September 28, 2026 to stop autonomous agents from circumventing application-layer controls to complete tasks. The platform pairs OpenShell, an open-source secure runtime that enforces agent boundaries on CPUs (with minimal overhead on NVIDIA Vera CPUs and extensibility to Arm and Intel platforms), with Sentry, an out-of-band watchdog running on NVIDIA BlueField-4 DPUs that continuously monitors agent behavior and can quarantine an agent that exceeds its boundaries in milliseconds. It is built on NVIDIA DOCA software for inspecting agent requests and responses, verifying identity, and enforcing zero-trust access policies independent of the software layer being governed. Over 100 organizations are collaborating on the platform, including Anthropic, Salesforce, SAP, Scale AI, and SpaceX. On October 21, 2026 at GTC Berlin, NVIDIA added the OpenShell Reference Architecture for Safe AI Agents, a full-stack design integrating OpenShell and Sentry across AI factories, RTX Spark and DGX Spark; a third-party recap names Anthropic, AWS, Google Cloud, Microsoft and OpenAI as partners, and says OpenShell is in its 0.1.0 series.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

  • OpenShell: open-source secure runtime enforcing agent boundaries on CPUs
  • Sentry: out-of-band watchdog on BlueField-4 DPUs that quarantines agents in milliseconds
  • Built on NVIDIA DOCA for request/response inspection, identity verification, and zero-trust policy enforcement
  • Hardware-based governance independent of the software layer, extensible to Arm and Intel CPUs
  • Full-stack governance across software, hardware, compute, and robotics systems
  • 100+ collaborating organizations including Anthropic, Salesforce, SAP, Scale AI, and SpaceX
  • OpenShell Reference Architecture (Oct 21, 2026) spanning AI factories, RTX Spark and DGX Spark

Use cases

  • Containing agents that attempt to bypass application-layer guardrails
  • Runtime quarantine of misbehaving autonomous agents in production
  • Zero-trust governance of agent fleets across mixed hardware vendors
  • Security auditing and policy enforcement for robotics and compute systems running agents

Pricing

OpenShell and related software available via NVIDIA developer resources and GitHub on a when-and-if-available basis; no pricing disclosed.

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released
28 Sept 2026
Record updated
22 Oct 2026
Record last checked
22 Oct 2026

Explore the Feed

Other entries filed under security and safety tooling; inclusion does not imply equivalent features.

  • DefenseClaw Edge Connector

    Cisco's open-source DefenseClaw gains an Edge Connector that allows or blocks AI agent tool calls directly on IoT and robot hardware.

  • Lakera Guard 2.0

    Lakera's real-time guardrail API for LLM apps, updated with multi-agent protection and a new benchmark for detecting indirect prompt injection in tool outputs.

  • Fleming-1

    Sierra's model that listens to a live phone call and scores whether the caller is an AI agent rather than a human.

Browse the Feed