SANTA CLARA, CALIFORNIA / RankWire.AI / – NVIDIA has launched the Open Agent Safety Platform, a comprehensive security solution designed for autonomous artificial intelligence agents. This platform integrates software-based controls with dedicated hardware monitoring to oversee AI workloads, covering stages from development and testing to deployment, including systems used in enterprise computing and robotics. NVIDIA emphasized that the platform can impose restrictions beyond the AI model itself, allowing organizations to choose individual components tailored to their computing and security needs.
OpenShell offers the platform’s open source runtime layer, responsible for managing how AI agents access digital resources. Administrators are able to define rules governing access to files, networks, tools, processes, and credentials. During task execution, the software logs agent activities and enforces these permissions. OpenShell establishes a security boundary around agents independently of instructions embedded within the models. Designed by NVIDIA to support both open-source and proprietary AI models, the software provides versatile security measures.
Adding a distinct monitoring layer, NVIDIA Sentry operates via the company’s BlueField-4 data processing units. The technology functions outside the main software environment of agents, continuously observing activity. When an agent’s actions surpass predefined security parameters, Sentry can isolate it within milliseconds. Furthermore, it can enforce access policies without relying on the agent to ensure compliance. NVIDIA’s DOCA software framework supports this system by facilitating identity verification, telemetry, inspection, and other security features.
Additional controls independently bolster agent security
This architecture separates the AI’s decision-making processes from the security enforcement systems. OpenShell confines agents within controlled runtime environments that incorporate isolation and policy enforcement. Meanwhile, Sentry continuously monitors these environments from dedicated hardware that remains unmanaged by the agents. NVIDIA developed the platform after security researchers documented instances where agents exceeded their intended application controls. The company explained that external enforcement mechanisms can grant operators direct control over what agents are allowed to access and execute.
According to NVIDIA, over 100 organizations are engaged with the technologies associated with 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. Their collaborations span cybersecurity, cloud infrastructure, enterprise software, and financial services sectors. Additionally, the platform supports autonomous systems operating within physical environments that necessitate controls beyond just application software.
The open source approach broadens hardware compatibility
OpenShell is compatible with NVIDIA Vera CPUs, which the company developed for demanding AI workloads involving agentic computing. Developers can also adapt the open source runtime to systems based on third-party architectures, with NVIDIA explicitly identifying Arm and Intel platforms as supported expansion options. Sentry depends on BlueField-4 hardware for its independent monitoring functions, effectively separating runtime controls from the hardware layer responsible for observing and restricting agent activity across various compatible computing environments.
NVIDIA founder and CEO Jensen Huang described AI safety and security as full-stack engineering challenges during the platform’s unveiling. The company presented the Open Agent Safety Platform as a reference architecture for securing agents across software, hardware, and infrastructure layers. OpenShell provides the software boundary, while Sentry delivers hardware-based independent oversight. Collectively, these components equip operators with tools to define permissions, log activities, and enforce restrictions as autonomous AI agents transition from development stages to operational environments.