How do Arm and NVIDIA secure AI agents that act on their own?

The trust boundary runs through silicon

AI agents do not just answer prompts. They run continuously. They call tools. They read data. They take actions to finish tasks. That persistence creates two distinct compute demands: one to execute the agent, another to police it. Arm supplies both sides. High‑performance CPUs host the agent workloads. Infrastructure processors such as DPUs sit outside the host and enforce policy. They isolate traffic. They observe behaviour without depending on the agent's own environment.

Arm‑based rack‑scale servers have overtaken x86 as the dominant platform for accelerated computing, according to IDC. Every major hyperscaler now deploys Arm technology.

The CPU ecosystem

The CPU ecosystem includes the Arm AGI CPU and AWS Graviton. Google Axion and Microsoft Cobalt are also part of it. NVIDIA's custom Vera CPU rounds out the list. Each is architecturally distinct yet compatible with the same Arm software stack.

OpenShell draws the runtime line

NVIDIA OpenShell establishes a secure runtime boundary around the agent. It governs what the agent can access and which actions it can take. Because OpenShell runs on the broad Arm CPU ecosystem, the same fundamental approach to secure agent execution extends across diverse cloud and AI infrastructure. The agent runtime depends on CPUs. The orchestration layer depends on CPUs. Tool execution and applications depend on CPUs. Those demands grow as agents become more capable and persistent. They also operate concurrently at scale.

BlueField provides the independent watchdog

Running the agent is only half the equation. Agentic infrastructure also needs independent compute that can observe and isolate. It must govern policies outside the agent's own environment. Arm‑based DPUs move networking and monitoring away from the host CPU. They also move isolation and security. They create a separate control point that remains independent of the agents and applications they protect.

BlueField‑4 demonstrates this model inside the Open Agent Safety Platform. The chip is powered by NVIDIA Grace with 64 Arm Neoverse V2 cores. It provides an independent infrastructure environment beyond the host. OpenShell integrates with NVIDIA Sentry on BlueField to extend policy enforcement from the agent runtime into that separate trust domain. Sentry acts as a real‑time watchdog. It monitors agent behaviour and can quarantine an agent exhibiting unsafe behaviour.

Security must follow the agent

Agentic AI will require more compute as agents become more capable and persistent, with growing autonomy. Security has to scale alongside that compute without forcing a trade‑off between the two. The Open Agent Safety Platform shows one way: Arm‑compatible compute runs the workloads while independent Arm‑based DPUs observe and isolate them, protecting them. As agents expand from cloud infrastructure to edge devices and physical systems, these security boundaries will need to follow wherever agents run and act. The document does not say how many agents are already deployed or what failure rates look like in production.