What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
NVIDIA’s September 28, 2026 announcement adds agent security to its AI infrastructure story: the Open Agent Safety Platform pairs OpenShell, a software runtime for setting boundaries around agents, with Sentry, an optional hardware-backed watchdog reference design. OpenShell does not require NVIDIA BlueField-4; Sentry’s described design does. NVIDIA has announced an architecture and integrations, not independent proof that the system prevents every agent escape.
What NVIDIA announced
NVIDIA describes the Open Agent Safety Platform as an open software platform and reference system design for governing and controlling agent software and the hardware and compute systems it uses. Its two components address different layers:
- OpenShell is an open-source runtime that sets software boundaries around agent execution and access on CPUs.
- Sentry is a reference system design for monitoring and responding to agent activity from a separate hardware domain, using NVIDIA BlueField-4 DPUs.
The practical distinction is that OpenShell supplies the software boundary; Sentry adds an out-of-band hardware layer. They are not the same component, and BlueField-4 is not a prerequisite for OpenShell.
How OpenShell is intended to secure an agent
NVIDIA says OpenShell can run agents in isolated sandboxes, restrict file and network access, filter system calls at the kernel level, and use a gateway for policy enforcement and auditing. It also describes policy verification as part of the design. These controls aim to limit what an agent can reach or do outside its assigned scope, rather than relying only on safeguards built into the model or application.
#1 Best Overall
NVIDIA says OpenShell can be extended to third-party compute platforms, including Arm and Intel. That is a compatibility claim from NVIDIA; the announcement materials do not establish the details or breadth of independently validated implementations.
How Sentry and BlueField-4 fit
Sentry is described as an out-of-band watchdog running on BlueField-4 DPUs. NVIDIA says it monitors agent activity independently of the agent’s software boundary and can quarantine or stop an agent that attempts to leave that boundary. The intended benefit is an enforcement and monitoring point outside the host software path being controlled.
This is an additional layer for deployments that want hardware-domain oversight, not a requirement for organizations using OpenShell by itself. NVIDIA’s description is architectural; it does not, by itself, demonstrate measured detection accuracy or prove that every escape attempt would be caught.
Choosing OpenShell alone or adding Sentry
| Consideration | OpenShell alone | OpenShell with Sentry |
|---|---|---|
| Enforcement location | Runtime and kernel-level software controls. | OpenShell controls plus an out-of-band hardware layer in NVIDIA’s Sentry reference design. |
| Hardware dependency | Does not require BlueField-4. | The described Sentry design uses BlueField-4 DPUs. |
| Use case emphasis | Setting access boundaries for agents in software. | Adding independent monitoring and enforcement where oversight outside the host software boundary is desired. |
| Evidence and operational details | The reviewed sources do not establish published performance measurements, detection accuracy, false-positive rates, deployment prerequisites, or incident-workflow details. Request these for the specific deployment and assess independent evaluations. | |
| Governance and interoperability | Check project governance and verify compatibility claims for the platforms and integrations you plan to use. | |
What adoption claims do—and do not—show
NVIDIA said on September 28, 2026, that more than 100 organizations were working with Open Agent Safety Platform technologies. It named organizations including Anthropic, Cisco, CrowdStrike, Dell Technologies, HPE, Hugging Face, JPMorganChase, Microsoft, Palo Alto Networks, Red Hat, Salesforce, SAP, and ServiceNow. The figure describes organizations working with the technologies; it is not a count of production deployments or customers.
Rank #3
NVIDIA also described specific collaborations and integrations: work with Anthropic around Claude Managed Agents, OpenShell, and BlueField; Salesforce integration of OpenShell with Slack for activity visibility and permission requests; and SAP embedding OpenShell with Joule Studio. These examples indicate ecosystem activity, but they are not independent evaluations of security outcomes.
What is still unknown
The announcement establishes NVIDIA’s design intent and describes how the components are meant to work. The reviewed sources do not provide independent customer benchmarks, measured detection accuracy, false-positive rates, or evidence that the platform prevents all agent escapes. Treat the capability descriptions as NVIDIA’s claims until independent evaluations or deployment evidence address those outcomes.
In an October 1, 2026 analysis, Network World’s Zeus Kerravala noted that NVIDIA had not published formal performance measurements at announcement. He also raised questions about neutral governance: NVIDIA said it planned to move OpenShell to the Linux Foundation, but that move had not occurred as of his analysis. These are dated observations and do not establish the project’s status after October 1.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What this means for AI infrastructure
NVIDIA’s pitch is that model-level safeguards and application-layer controls may not be enough when agents act across files, tools, networks, and other systems. OpenShell and Sentry extend the security discussion into the runtime and infrastructure layers. That is a meaningful shift in emphasis for a company associated with AI compute, but the practical value will depend on implementation, operational fit, and evidence of effectiveness—not the announcement alone.
Free tools Windows power users keep installed
One-click scans. No signup required.
Best Value
As NVIDIA CEO Jensen Huang put it in the September 28 release, “Safety and security require full-stack engineering.” The platform reflects that premise by placing controls around agent execution as well as within software. Whether those controls deliver reliable protection remains a question for measured, independently scrutinized deployments.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
