Summary
NVIDIA announced an open platform combining OpenShell, software that sets runtime boundaries for AI agents, with Sentry, a hardware watchdog reference design for BlueField-4 DPUs. NVIDIA says OpenShell is broadly available; Sentry is described as part of the reference design.
NVIDIA announced its Open Agent Safety Platform on September 28, 2026, combining software controls for AI agents with a hardware-based monitoring design. The platform’s two named components are OpenShell, an open-source secure runtime that sets boundaries for agent actions, and Sentry, a watchdog reference design that runs on NVIDIA BlueField-4 data-processing units (DPUs).
An AI agent can use a model and connected tools to carry out tasks across multiple steps. NVIDIA’s approach places controls around that activity: OpenShell governs what an agent can do in its runtime, while Sentry is designed to monitor activity from a separate, out-of-band position. NVIDIA presents the platform as a way for organisations to apply governance across software, computing infrastructure and, where relevant, robotics systems.
Two layers of agent control
OpenShell traces agent actions and enforces policies as agents run on CPUs. NVIDIA says it provides a boundary outside the model and agent harness, allowing organisations to control how agents execute tasks across both open and closed models. The company says OpenShell is broadly available and can run on NVIDIA Vera CPUs, which it describes as purpose-built for agentic AI. As open-source software, OpenShell can also be extended to work with third-party compute platforms, including those from Arm and Intel.
Sentry adds a separate monitoring layer to NVIDIA’s reference system design. It runs on BlueField-4 DPUs and uses an isolated, out-of-band trust domain to inspect agent activity and enforce security policies independently in silicon. NVIDIA says Sentry can quarantine and stop an agent in milliseconds if it tries to move outside its software boundary.
Sentry is built on NVIDIA DOCA software. The company says DOCA provides capabilities for inspecting agent requests and responses, producing attested telemetry, verifying agent identity and applying granular access policies to data, tools, application programming interfaces and services. In practical terms, the design pairs rules that constrain an agent’s runtime with a separate hardware-based mechanism intended to detect and contain boundary violations.
Integrations and availability
The announcement describes the platform as a collection of components that organisations can deploy according to their requirements, rather than a single mandatory configuration. NVIDIA says OpenShell and related software, including skills, are available through its developer resources and GitHub. Sentry is presented as part of a reference system design; the announcement does not describe it as a separately available product.
Several organisations are working with NVIDIA on integrations. Anthropic says its Claude Managed Agents run the agent loop separately from the sandboxes where work executes, with OpenShell and BlueField integrations intended to give enterprises additional control over sandbox access. Salesforce and NVIDIA have integrated OpenShell with Slack so teams can view agent activity and audit events, and approve or reject requests for additional permissions. SAP is embedding OpenShell in Joule Studio runtime and contributing engineering work to the software.
NVIDIA says more than 100 organisations are working with its platform technologies, including companies in software, finance, energy and robotics. These integrations indicate the intended range of uses, but the release also notes that some described products and features are at different stages and may be offered on a when-and-if-available basis, with timing subject to change.