SANTA CLARA, CALIFORNIA / RankWire.AI / – NVIDIA has launched the Open Agent Safety Platform, an advanced security solution designed specifically for autonomous AI agents. By integrating software controls with hardware-based monitoring, the platform ensures comprehensive oversight across AI operations. It is applicable throughout development, testing, and deployment phases, including enterprise computing and robotics systems. NVIDIA highlighted that the platform can implement restrictions beyond the AI model itself, allowing organizations to tailor components based on their specific security and computing needs.
OpenShell offers an open-source runtime environment that manages how AI agents access digital resources. Administrators can establish rules regulating access to files, networks, tools, processes, and credentials. The software not only logs agent activities but also enforces permissions during task execution. This setup creates a security boundary around agents without embedding instructions inside the models. NVIDIA designed OpenShell to support both open and proprietary AI models, ensuring flexibility for various implementations.
NVIDIA Sentry adds an additional layer of security through the company’s BlueField-4 data processing units. The technology runs independently outside the main AI software environment and continuously monitors agent activity. It can isolate an agent within milliseconds if its behavior exceeds predefined security parameters. Furthermore, Sentry enforces access policies without relying on the agent to ensure compliance. NVIDIA utilizes its DOCA software framework to facilitate identity verification, telemetry, inspection, and other security functions.
Separate controls enhance AI agent security
This architecture distinctly separates the decision-making process of AI from the systems tasked with enforcing security rules. OpenShell encapsulates agents within controlled runtime environments, ensuring isolation and policy enforcement. Meanwhile, Sentry monitors these environments through dedicated hardware that remains unmanaged by the AI agents. The platform was developed in response to security researchers documenting incidents where agents bypassed intended application controls. NVIDIA asserts that external enforcement provides operators with direct control over what agents can access and execute.
More than 100 organizations are now working with technologies related to the Open Agent Safety Platform. Participants include Anthropic, Microsoft, Cisco, CrowdStrike, Dell Technologies, HPE, and Hugging Face. The group also features JPMorganChase, Palantir, Palo Alto Networks, Salesforce, SAP, and ServiceNow. These entities operate across sectors such as cybersecurity, cloud infrastructure, enterprise software, and financial services. The platform additionally supports autonomous systems interacting with physical environments that require controls beyond standard application software.
Open source approach broadens hardware compatibility
OpenShell is compatible with NVIDIA Vera CPUs, which are designed for demanding AI workloads involving agents. Developers also have the option to extend the open-source runtime to third-party architectures. NVIDIA specifically identified support for Arm and Intel platforms as part of its expansion plans. Sentry relies on BlueField-4 hardware for its monitoring role, separating runtime controls from the hardware layer responsible for observing and restricting agent activity across various computing environments.
Jensen Huang, NVIDIA’s founder and CEO, emphasized that AI safety and security are comprehensive engineering challenges when unveiling the platform. He described the Open Agent Safety Platform as a reference system to secure agents across software, hardware, and infrastructure layers. OpenShell offers the software boundary, while Sentry provides independent hardware oversight. This combination equips operators with the tools to set permissions, log activities, and enforce restrictions as autonomous AI agents transition from development to operational deployment.