SANTA CLARA, CALIFORNIA / RankWire.AI / – NVIDIA has announced the launch of the Open Agent Safety Platform, a pioneering security solution designed for autonomous artificial intelligence agents. This system integrates software controls with dedicated hardware surveillance across various AI workloads, encompassing development, testing, and deployment phases. It supports systems utilized in enterprise computing and robotics. NVIDIA emphasized that the platform is capable of implementing restrictions beyond the AI model itself, allowing organizations to choose individual components tailored to their specific computing and security needs.
Back in the early stages of deployment, OpenShell serves as the platform’s open-source runtime layer, managing how AI agents access digital resources. Administrators have the ability to establish rules concerning file, network, tool, process, and credential access. During operation, the software logs agent activities and enforces these permissions. OpenShell constructs a security boundary around agents without relying on embedded instructions within the models. NVIDIA engineered this software to be compatible with both open-source and proprietary AI models.
In recent developments, NVIDIA Sentry adds a separate layer of monitoring through the company’s BlueField-4 data processing units. The technology functions outside an agent’s primary software environment, continuously observing its activity. When an agent’s actions exceed predefined security limits, Sentry can isolate it within milliseconds. Moreover, it can enforce access policies independently of the agent, enhancing security. NVIDIA utilizes its DOCA software framework to facilitate identity verification, telemetry, inspection, and various security measures.
Decoupled controls enhance the security of AI agents
In the initial stages of the architecture’s development, the separation of AI decision-making from security enforcement systems was emphasized. OpenShell ensures agents operate within controlled runtime environments, providing isolation and policy application. Meanwhile, Sentry monitors these environments from dedicated hardware that is not managed by the agent. The platform was developed in response to documented cases where agents moved beyond their intended application controls, as noted by security researchers. NVIDIA noted that external enforcement offers operators direct oversight of agent access and execution capabilities.
As of now, NVIDIA reports over 100 organizations engaged with technologies related to the Open Agent Safety Platform. The list includes Anthropic, Microsoft, Cisco, CrowdStrike, Dell Technologies, HPE, and Hugging Face. Additional participants are JPMorganChase, Palantir, Palo Alto Networks, Salesforce, SAP, and ServiceNow. These entities operate across sectors such as cybersecurity, cloud infrastructure, enterprise software, and financial services. Furthermore, the platform supports autonomous systems interacting with physical environments, where additional controls are necessary beyond software alone.
Initially, OpenShell is compatible with NVIDIA Vera CPUs, specially designed for high-demand agentic AI workloads. Developers also have the option to extend the open-source runtime to systems based on third-party architectures, with support explicitly identified for Arm and Intel platforms. Sentry relies on BlueField-4 hardware for its independent surveillance function, effectively separating runtime controls from the hardware layer that monitors and restricts agent activity across compatible computing environments.
During the platform’s announcement, NVIDIA founder and CEO Jensen Huang highlighted AI safety and security as comprehensive full-stack engineering challenges. He positioned the Open Agent Safety Platform as a reference model for securing agents across both software and hardware layers. OpenShell supplies the software boundary, while Sentry offers independent hardware oversight. The combined setup equips operators with tools to define permissions, log activity, and enforce restrictions as autonomous AI agents transition from development stages into operational environments.