The Pulse
NVIDIA Launches Open Agent Safety Platform With Runtime Controls
NVIDIA announced its Open Agent Safety Platform on September 28, combining the OpenShell agent runtime with a BlueField-based monitoring design called Sentry. The system, the company says, can enforce policies and quarantine agents that cro
AI.info Team · Written with AI · fact-checked ·
On September 28, NVIDIA announced an open agent-safety platform that pairs software controls around an agent with a separate hardware monitor designed to detect when the agent crosses its boundaries. The system combines OpenShell, NVIDIA’s secure runtime, with Sentry, a reference design that monitors agent activity using BlueField-4 data processing units.
OpenShell traces agent actions and enforces policies, while Sentry watches from a separate, “out-of-band” domain. If an agent tries to move beyond its software boundary, the company says, Sentry can quarantine and stop it in milliseconds. The release describes the platform as a way to govern agents across software, computing hardware and robotics systems.
OpenShell sets the rules; Sentry watches outside them
OpenShell is intended to establish a runtime boundary beyond the model and the agent harness—the software that coordinates an agent’s actions and tools. The software is now broadly available and can control autonomous agents running with open or closed models. OpenShell runs on NVIDIA Vera CPUs, but the open-source software can also be extended to third-party platforms, including Arm and Intel systems, the company says.
Sentry adds a second layer that operates independently of the agent. Running on BlueField-4 DPUs, it monitors activity and enforces policies from an isolated trust domain. The design can inspect requests and responses, verify agent identity and apply access rules to data, tools, application programming interfaces and services.
The distinction matters to NVIDIA’s proposal: an agent’s own software controls may not be enough if an agent can circumvent them. A monitor outside that software boundary is meant to provide a separate point of enforcement. Those capabilities and response times are NVIDIA’s descriptions of its platform, not evidence in the announcement of independent testing.
More than 100 organizations are working with the technology
More than 100 organizations are working with NVIDIA’s platform technologies, including Anthropic, Cisco, CrowdStrike, Hugging Face, Microsoft, Palantir, Salesforce, SAP and ServiceNow. The announcement describes several different forms of participation rather than a single deployment model. Anthropic, for example, is integrating OpenShell and BlueField with Claude Managed Agents, while Salesforce and NVIDIA have integrated OpenShell with Slack so teams can review agent activity and approve or reject requests for additional permissions.
Other examples span enterprise software and infrastructure. Red Hat is running OpenShell and NVIDIA’s DOCA software on Red Hat AI Factory with NVIDIA; robotics companies including Figure, Gecko Robotics and Skild AI are building with OpenShell for autonomous systems. The platform’s software and skills are available through NVIDIA’s developer resources and GitHub.
NVIDIA frames agent security as a system problem
Jensen Huang, NVIDIA’s founder and CEO, tied the launch to a broader call for controls beyond the model itself. “Safety and security require full-stack engineering,” he said, describing the platform as a way for companies, researchers and public-sector organizations to share practices and align on evaluation methods.
The announcement positions OpenShell and Sentry as parts of a larger security effort, not a guarantee that agents will behave safely in every setting. Organizations can deploy platform components according to their needs, and NVIDIA’s release notes that some described products and features remain subject to change. The concrete offer today is OpenShell software and a hardware-monitoring reference design, with partners developing integrations for their own agent systems.