NVIDIA Open Agent Safety Platform Integrations

NVIDIA OpenShell is an open and secure runtime environment tailored for the operation of autonomous AI agents, managing their execution, access privileges, and the routes for inference traffic. Rather than incorporating security measures directly into the model or application, it upholds safety protocols in the surrounding infrastructure, ensuring that nothing is permitted by default; permissions are allocated through predefined policies, with enforcement mechanisms acting outside the agent's process to avert any potential bypass through prompts. Each autonomous agent is confined within a controlled sandbox, which limits direct network access, restricts file visibility, and employs kernel-level monitoring of system calls to guarantee stringent oversight. Serving as the control plane, a gateway handles user authentication, manages the lifecycle of sandboxes, and provides the agents with policies, configurations, credentials, and settings for inference. Furthermore, an external supervisor evaluates network requests based on policies at multiple levels—including binary, destination, method, and path—and grants access to credentials only when expressly authorized, thereby bolstering the overall security architecture. This comprehensive and layered framework ensures that the agents operate efficiently while upholding strict security measures throughout their processes, ultimately fostering a trustworthy environment for AI interaction. The attention to detail in security design reflects a commitment to safeguarding both the agents and the systems they interact with.

NVIDIA DOCA is a robust platform designed to maximize the capabilities of NVIDIA BlueField DPUs and NVIDIA ConnectX SuperNICs, particularly for agentic AI environments. It presents a comprehensive collection of SDKs, libraries, production-ready microservices, drivers, management utilities, and orchestration frameworks that streamline the development and oversight of accelerated infrastructure services, which include networking, AI-focused storage, cybersecurity, observability, and lifecycle management. By leveraging DOCA, developers can effectively offload, accelerate, and isolate infrastructure tasks, allowing host computing resources to be freed up for a variety of AI services and applications. The networking features encompass virtual switching, packet processing, multi-tenant isolation, sophisticated traffic management, RDMA acceleration, and software-defined networking functionalities. In the realm of storage, DOCA is equipped to handle KV cache infrastructure, NVMe over Fabrics, storage virtualization, and software-defined storage services, promoting a seamless integration of various storage solutions. Furthermore, the utilization of DOCA not only boosts operational efficiency but also enhances the flexibility required for managing sophisticated infrastructure. This adaptability is crucial as organizations increasingly rely on advanced technology to meet their dynamic operational demands.