Arm, Nvidia Build Agentic AI Security Infrastructure
TECHWORLD ·
✦ AI Summary
As agentic AI spreads, computing infrastructure is increasingly required to deliver both performance and security.
Arm announced on the 2nd that it will work with Nvidia to build secure computing infrastructure, outlining high-performance CPUs and DPUs with independent security and control functions as the two support areas.
Nvidia unveiled Open Agent Safety Platform and implemented a structure that separates agent execution from monitoring and isolation through BlueField-4 and OpenShell and Sentry.
As agentic AI moves beyond question-and-answer interactions into a stage where it can plan on its own and use tools to execute tasks, the role of computing infrastructure is expanding as well. Because agents run continuously and access data and systems, infrastructure now needs not only workload processing performance but also security mechanisms that control agent behavior and functions that block potential threats.
Related to this, Arm announced on the 2nd that it will work with Nvidia to build a secure computing infrastructure for the age of agentic AI. Arm outlined two support areas in response to the spread of agentic AI.
The first is a high-performance CPU for running agents. The second is a DPU that provides independent security and control functions.
Nvidia unveiled Open Agent Safety Platform. The platform was introduced as an implementation example of the architecture Arm described.
Arm supplies AI and cloud infrastructure technologies to major hyperscalers. According to IDC, Arm-based rack-scale servers surpassed x86 in the server market for accelerated computing. As a result, Arm-based rack-scale servers have become the mainstream platform in the accelerated computing server market, and the shift to Arm in AI infrastructure is also expanding.
As this transition accelerates and agentic AI spreads, the importance of the Arm-based computing ecosystem is also growing. In agentic AI infrastructure, the CPU handles the agent runtime, orchestration, tool execution, and application operations. In addition, as AI agents become more capable, CPU demand rises; as AI agents run for longer periods, CPU demand rises; and as the scale of concurrent AI agent processing expands, CPU demand rises.
Arm has its own Arm AGI CPU. Arm is also supporting agentic AI and cloud workloads through the AWS Graviton-based CPU ecosystem, the Google Axion-based CPU ecosystem, the Microsoft Cobalt-based CPU ecosystem, and the Nvidia Vera-based CPU ecosystem. Arm said it is supporting agentic AI and cloud workloads through these diverse Arm-based CPU ecosystems.
Nvidia's OpenShell is a technology that supports the secure execution of agents in the Arm CPU ecosystem. OpenShell establishes a secure runtime boundary around agents, manages what agents can access, and manages what tasks agents can perform. Through this, OpenShell enables secure execution in Arm CPU environments by managing agent boundaries and privileges.
The effect of this technology lies in building an agent execution environment that is not dependent on a specific CPU, and it can be applied to various Arm-based cloud and AI infrastructures. Accordingly, OpenShell provides a consistent agent execution environment across a range of Arm-based cloud and AI infrastructures. Against the backdrop of expanding use cases for agentic AI, this has the significance of laying the groundwork for applying security policies and ensuring secure execution across diverse computing environments.
However, securing only the agent execution environment makes it difficult to fully secure agentic AI infrastructure. Accordingly, an additional control layer separate from the execution environment is required. This separate control layer observes the system, applies security policies, and isolates threats when necessary.
This structure works through Arm-based DPUs and infrastructure processors, separating networking, monitoring, isolation, and security functions from the host CPU and establishing an independent control point outside the agent and application execution environment. Accordingly, infrastructure and network control functions operate separately from the agent itself, allowing security policies at the infrastructure level to be managed regardless of the agent's execution status.
Nvidia BlueField implements this structure in Open Agent Safety Platform. BlueField-4 is equipped with 64 Arm Neoverse V2 cores and is powered by Nvidia Grace, providing an infrastructure environment that operates independently of the host system. Nvidia OpenShell and Sentry are also integrated into BlueField-4, extending the scope of policy enforcement beyond the agent execution environment to a separate trust domain.
OpenShell controls access and tasks within the agent execution environment. In contrast, Nvidia Sentry takes on the role of strengthening system-level security through an independent monitoring layer. Sentry monitors agent behavior in real time and helps isolate agents that exhibit unsafe behavior.
This structure focuses on separating agent execution functions from security management functions. The goal is to implement a broader system-level security model. This division of roles is realized through a model in which Arm CPUs run agent workloads and Arm-based DPUs independently observe, manage, and isolate agents.
The challenges raised by the development of agentic AI are the concurrent expansion of computing infrastructure performance and security systems. As agents become more autonomous and persistent, more computing resources are needed. At the same time, security functions must also be strengthened to protect the data and systems agents access.
Nvidia presented Open Agent Safety Platform. The platform's approach is to run agents in an Arm-compatible computing environment and provide monitoring and isolation functions through an independent Arm-based DPU. This security architecture proposed by Nvidia is not limited to cloud data centers. As agentic AI expands to edge devices and physical AI systems as well, the importance of applying security boundaries across diverse execution and operating environments is increasing.
In response, Arm plans to expand its computing ecosystem across AI infrastructure. The scope of this expansion encompasses CPU technology and DPU technology. Arm's plan is focused on supporting the performance and security of agentic AI. Arm said it envisions building a computing architecture that connects agent execution, behavior control, and system protection to support the secure spread of agentic AI.
Source: TECHWORLD · Park Kyu-chan
Original: https://www.epnc.co.kr/news/articleView.html?idxno=407714
References
This article was produced with the help of an automated content generation algorithm.
Source: TECHWORLD
View originalThis article was summarized and organized by BizCrush based on the original article from TECHWORLD. For exact quotations and full details, please refer to the original article.