SANTA CLARA, CALIFORNIA / RankWire.AI / – NVIDIA has unveiled its Open Agent Safety Platform to bolster security measures for autonomous AI agents. Covering agent testing, deployment, software, computing infrastructure, and robotics systems, the platform was announced by NVIDIA on September 28 as enterprises increase their deployment of agents across complex computing environments. The system integrates open source runtime software with independent hardware monitoring, allowing organizations to customize deployment of platform components based on their security and operational needs.
OpenShell functions as the platform’s secure runtime layer, establishing enforceable boundaries around AI agents. This open source software traces agent activities and enforces policies while agents operate on CPUs. Operators can set access controls for files, networks, tools, processes, and credentials prior to agent activation, with OpenShell ensuring these limits are maintained during operation. NVIDIA stated that the software supports both open and closed AI models and creates a security boundary outside the model and its agent harness.
The other key element, NVIDIA Sentry, offers independent monitoring via BlueField-4 data processing units. Operating outside the agent’s software environment, Sentry continuously observes agent behavior and can quarantine an agent within milliseconds if it attempts to breach its permitted boundaries. It enforces security policies independently through hardware. Built on NVIDIA’s DOCA software, Sentry supports agent identity verification, access controls, telemetry, and inspection of requests and responses.
Runtime controls extend beyond the model
This platform links application software, secure runtime controls, and computing infrastructure into a layered security architecture. OpenShell isolates agents within sandboxed environments with kernel-level separation and policy enforcement, while Sentry adds an additional hardware barrier that agents cannot directly manipulate. Recent security breaches have demonstrated how agents can bypass application-level controls while performing tasks, prompting NVIDIA to design a system that provides organizations with controls that are separate from the AI agent and the model guiding its actions.
Over 100 organizations are collaborating with technologies related to the Open Agent Safety Platform, according to NVIDIA. Anthropic has worked with NVIDIA to develop additional controls for managed AI agents and sandbox environments. Salesforce has integrated OpenShell with Slack to enable teams to review agent activities, audit events, and permission requests. Participants also include firms from financial services, enterprise software, cybersecurity, cloud infrastructure, robotics, and energy sectors. The platform additionally supports security controls for autonomous systems functioning in physical environments.
Open source software broadens deployment options
OpenShell operates alongside NVIDIA Vera, the company’s CPU optimized for agentic AI workloads. NVIDIA also permits developers to adapt this open source software for third-party platforms, including systems utilizing Arm and Intel architectures. The entire platform can run across compatible hardware environments, while Sentry relies on BlueField-4 for its independent monitoring capabilities. NVIDIA stated that the Open Agent Safety Platform software, including OpenShell and related components, is accessible through its developer resources and the GitHub platform.
NVIDIA founder and CEO Jensen Huang emphasized that AI safety and security are comprehensive engineering challenges, and the company contributes technology to the Open Secure AI Alliance, an initiative managed by the Linux Foundation. This alliance promotes open research, tools, and security practices for AI agents. NVIDIA highlighted that its platform supports governance across software, hardware, computing, and robotics layers. The release introduces new runtime and hardware controls as organizations expand their use of autonomous AI agents in diverse business systems.