SANTA CLARA, CALIFORNIA / RankWire.AI / – NVIDIA has introduced its Open Agent Safety Platform to enhance security measures for autonomous AI agents. This framework encompasses testing, deployment, software, infrastructure, and robotics systems. NVIDIA revealed the technology on September 28, as enterprises broaden their deployment of agents within intricate computing environments. The system integrates open source runtime software alongside independent hardware monitoring. NVIDIA indicated that organizations can customize the deployment of individual platform modules based on their security and operational needs.
OpenShell functions as the platform’s secure runtime layer, establishing enforceable boundaries around AI agents. The open source software monitors agent activity and enforces policies during execution on CPUs. Administrators have the ability to set permissions for file access, network connections, tools, processes, and credentials prior to agent operation. During runtime, OpenShell ensures these restrictions are upheld. NVIDIA noted that the software is compatible with both open and closed AI models, providing a security boundary outside the model and its agent harness.
The other key component, NVIDIA Sentry, offers autonomous monitoring through BlueField-4 data processing units. Sentry functions independently of the agent’s software environment, continuously observing agent behavior. The system can isolate an agent within milliseconds if it attempts to exceed its authorized boundaries. Furthermore, security policies are enforced independently through dedicated hardware. NVIDIA developed Sentry utilizing its DOCA software, which facilitates agent identity verification, access control, telemetry, and inspection of agent requests and responses.
Runtime controls now extend beyond the AI model itself
This platform links application software, secure runtime controls, and computing infrastructure into a multi-layered security architecture. OpenShell isolates agents within sandboxed environments featuring kernel-level separation and policy enforcement. Sentry introduces a distinct hardware layer that agents cannot manipulate directly. NVIDIA highlighted recent security breaches where agents bypassed application-level controls to complete tasks. The system was designed to give organizations controls that remain separate from both the AI agent and the underlying model guiding its actions.
Over 100 organizations are currently collaborating with technologies related to the Open Agent Safety Platform, according to NVIDIA. Anthropic has partnered with NVIDIA to develop additional controls for managed AI agents and sandbox environments. Salesforce has integrated OpenShell with Slack, enabling teams to review agent activities, audit events, and permission requests. The participant base extends across sectors such as financial services, enterprise software, cybersecurity, cloud infrastructure, robotics, and energy. The platform also offers security controls tailored for autonomous systems operating within physical environments.
Open source technology broadens deployment options
OpenShell runs on NVIDIA Vera, the company’s CPU optimized for agentic AI workloads. Developers are also allowed to extend this open source software to third-party hardware platforms, including those utilizing Arm and Intel architectures. The comprehensive platform can operate across various compatible hardware environments, with Sentry leveraging BlueField-4 for independent monitoring. NVIDIA stated that the Open Agent Safety Platform software, along with associated skills and tools, is accessible through its developer resources and the GitHub platform.
NVIDIA founder and CEO Jensen Huang emphasized that AI safety and security are full-stack engineering challenges. Additionally, the platform contributes to the Open Secure AI Alliance, an initiative overseen by the Linux Foundation. This alliance emphasizes open research, security practices, and tools for AI agents. NVIDIA confirmed that its platform supports governance spanning software, hardware, computing, and robotics layers. The release introduces new runtime and hardware controls as more organizations implement autonomous AI agents across an expanding array of business systems.
}#eilxd3ky5n6#