SANTA CLARA, CALIFORNIA / RankWire.AI / – NVIDIA has unveiled the Open Agent Safety Platform, a new security framework aimed at safeguarding autonomous artificial intelligence agents. This innovative system integrates software controls with dedicated hardware oversight across various AI tasks, spanning development, testing, and deployment phases, including systems utilized in enterprise computing and robotics. NVIDIA emphasized that the platform can impose restrictions beyond the confines of the AI model itself. Additionally, organizations can tailor the selection of individual components based on their specific computing and security needs.
OpenShell serves as the platform’s open-source runtime layer, managing how AI agents access digital resources. It allows administrators to establish rules governing access to files, networks, tools, processes, and credentials. The software tracks agent activities and enforces permissions during task execution. OpenShell establishes a security perimeter around agents without relying on embedded instructions within the model. NVIDIA designed this component to be compatible with both open and proprietary AI models.
NVIDIA Sentry adds an additional monitoring layer via the company’s BlueField-4 data processing units. The technology functions outside the main software environment of an agent, enabling continuous activity observation. Sentry can isolate an agent within milliseconds if its actions breach predefined security thresholds. Furthermore, it can enforce access policies independently of the agent’s compliance. NVIDIA leverages its DOCA software framework to facilitate identity verification, telemetry, inspection, and other security functionalities.
Independent controls bolster the security of AI agents
This architecture distinctly separates AI decision-making processes from the systems responsible for enforcing security protocols. OpenShell places agents within controlled runtime environments that ensure isolation and policy enforcement. Meanwhile, Sentry monitors these environments through dedicated hardware that the agent itself does not manage. The platform’s development was prompted by security researchers who identified incidents where agents operated beyond their intended application controls. NVIDIA highlighted that external enforcement mechanisms provide operators with direct oversight over what agents can access and execute.
According to NVIDIA, over 100 organizations are engaged with technologies related to the Open Agent Safety Platform. Participating entities include Anthropic, Microsoft, Cisco, CrowdStrike, Dell Technologies, HPE, and Hugging Face. The collaboration also features JPMorganChase, Palantir, Palo Alto Networks, Salesforce, SAP, and ServiceNow. Their collective efforts span cybersecurity, cloud infrastructure, enterprise software, and financial services. The platform is also designed to support autonomous systems interacting with physical environments, requiring controls beyond the software layer.
The open-source approach expands hardware compatibility
OpenShell operates on NVIDIA Vera CPUs, which are engineered for intensive agentic AI computing workloads. Developers can extend the open-source runtime to systems built on third-party architectures, with NVIDIA specifically identifying Arm and Intel platforms as supported options. Sentry depends on BlueField-4 hardware to fulfill its independent monitoring function. This configuration separates runtime control from the hardware layer responsible for observing and restricting agent activity across compatible computing environments.
Jensen Huang, NVIDIA’s founder and CEO, described AI safety and security as comprehensive full-stack engineering challenges during the platform’s announcement. The company positioned the Open Agent Safety Platform as a reference framework for securing agents throughout software, hardware, and infrastructure layers. OpenShell provides the software boundary, while Sentry offers hardware-based independent oversight. Together, these components enable operators to define permissions, log activities, and enforce restrictions as autonomous AI agents transition from development to real-world operational environments.