NVIDIA launched its Open Agent Safety Platform on 28 September 2026, offering controls for autonomous AI agents from testing through deployment. The platform combines OpenShell, an open-source runtime that traces agent activity and enforces policies, with Sentry, an optional hardware-based watchdog using NVIDIA BlueField-4 data processing units (DPUs). NVIDIA says Sentry can independently monitor agents and quarantine or stop them within milliseconds if they move beyond permitted boundaries.
OpenShell creates a runtime security boundary around agents running on CPUs, controlling access to systems, data and services while recording actions. It combines sandboxed execution, controlled service access, credential management and kernel-level controls over files and processes. A supervisor checks outbound requests against policy, while human reviewers can approve or reject requests for extra permissions; Salesforce has integrated this capability with Slack for activity review and auditing. NVIDIA says OpenShell is broadly available as open source and can be extended to Arm and Intel compute platforms.
Sentry operates from an isolated trust domain outside the agent’s software environment, applying governance controls to requests involving data, tools, application programming interfaces (APIs) and services. More than 100 organisations, including Anthropic, Microsoft, CrowdStrike, Palo Alto Networks, SAP, Salesforce and ServiceNow, are working with the platform, according to NVIDIA. The company said it is also being explored in financial services, energy and robotics, including systems that can affect the physical world.