Quick answer
NVIDIA announced the Open Agent Safety Platform on September 28, 2026. It combines OpenShell, an open-source runtime for setting and enforcing an agent's software permissions, with Sentry, a reference design for hardware-based monitoring on BlueField-4 DPUs. OpenShell is available through NVIDIA's developer resources and GitHub; Sentry's out-of-band enforcement depends on specific NVIDIA hardware. These are infrastructure controls for agent builders, not a guarantee that an agent is safe or a feature announcement for Chat AI.
OpenShell sets a software boundary around an agent
OpenShell is the software runtime in NVIDIA's platform. NVIDIA describes it as a way to sandbox agents, trace their actions, and enforce policy while they work, including when they use open or closed models. The public GitHub project provides a quickstart and lists Linux, macOS on Apple Silicon, and experimental Windows support through WSL 2, with a container or host-virtualization option. That is project documentation, not a hands-on test for this article.
Sentry is a separate hardware monitoring layer
Sentry is NVIDIA's reference design for an out-of-band watchdog running on BlueField-4 data processing units. NVIDIA says it monitors agent behavior independently of the software runtime and can quarantine an agent that crosses its boundaries. This part of the design depends on BlueField-4 hardware; it is not simply another software package to install beside OpenShell on any computer.
Sources: NVIDIA, NVIDIA Developer
The practical distinction is software policy versus independent enforcement
For a developer evaluating the announcement, OpenShell is the immediately inspectable runtime and policy surface; Sentry is the hardware-backed reference layer NVIDIA describes for monitoring and intervention outside that runtime. NVIDIA says OpenShell is open source and can be extended for third-party compute platforms, while its platform is optimized for Vera CPUs and BlueField-4 systems. Check the current support matrix and documentation before choosing a deployment path.
Sources: NVIDIA, NVIDIA Developer, NVIDIA
Guardrails still depend on the permissions and policies people configure
The announcement describes controls for limiting what agents can access and how they act; it does not establish that every harmful action can be prevented. The Associated Press also quotes an outside computer-security researcher noting that choosing a useful, least-privilege policy can be difficult. Treat NVIDIA's performance and protection descriptions as vendor claims, not independent test results, and review the policy, tools, credentials, and monitoring in any real deployment.
Sources: NVIDIA, Associated Press
This is agent infrastructure, not a new consumer chatbot feature
Open Agent Safety Platform is aimed at people building or operating agent systems. It does not add a model to Chat AI or change which models the app offers. Readers comparing everyday AI assistants can use Chat AI to access its listed models in one app; this NVIDIA announcement concerns infrastructure controls rather than that consumer workflow.
Sources: NVIDIA
Frequently asked questions
What readers usually ask
What is NVIDIA Open Agent Safety Platform?
It is NVIDIA's agent-security platform combining OpenShell runtime software with Sentry, a hardware-based reference monitoring design for BlueField-4 DPUs.
What is the difference between OpenShell and Sentry?
OpenShell creates and enforces a software sandbox and policy boundary around an agent. Sentry is a separate out-of-band hardware layer that NVIDIA says can monitor behavior and quarantine agents that leave those boundaries.
Can I try OpenShell without NVIDIA hardware?
NVIDIA says the open-source OpenShell software can be extended to third-party compute platforms, including Arm and Intel. Its GitHub quickstart lists Linux, macOS on Apple Silicon, and experimental Windows WSL 2 support. Sentry's announced reference design uses BlueField-4 DPUs. Check the current project documentation for requirements.
Does NVIDIA's platform guarantee that an AI agent is safe?
No. It describes runtime and hardware controls, not a guarantee against every failure. The permissions and policies still need to be designed, reviewed, and monitored for the system's use case.
Does this announcement add an AI model or feature to Chat AI?
No. NVIDIA announced developer and infrastructure tools; it did not announce a Chat AI model or app feature.
Evidence
Sources
- NVIDIA launches Open Agent Safety Platform to secure agents from testing to deploymentNVIDIA · Primary source
- NVIDIA Open Agent Safety Platform: a reference for continuous in-silicon agent monitoringNVIDIA Developer · Primary source
- NVIDIA OpenShell project and quickstartNVIDIA · Primary source
- Nvidia unveils security platform to stop AI agents from going rogueAssociated Press · Secondary source