NVIDIA DOCA: embedded security for AI agent infrastructure

NVIDIA published a piece in its corporate developer blog (NVIDIA Developer Blog) about how the NVIDIA DOCA platform provides built-in security for the infrastructure on which autonomous AI agents operate. The starting point of the article is the thesis that the era of AI is forming a new class of infrastructure: the so-called "AI factories" (AI factories), which convert data into intelligence for autonomous agents acting with unprecedented scale of tasks and level of independence.

What DOCA is and why AI infrastructure needs a separate security layer

• Platform — NVIDIA DOCA, a software framework for DPU (data processing unit)
• Target equipment — cards from the NVIDIA BlueField family
• NVIDIA concept — "AI factories" (AI factories), data centers producing intelligence as a product
• Target workload — autonomous AI agents with delegated authority
• Source — NVIDIA developer blog (NVIDIA Developer Blog)

NVIDIA DOCA is the company's software framework for DPU, BlueField class devices that take on part of a data center's network and infrastructure functions, offloading core computing resources from CPU and GPU. The idea behind the platform is to move control over traffic, access, and security policies to a separate level isolated from the main computing workload — so that compromising an application or agent on one server does not automatically open access to the rest of the data center infrastructure. This principle of layer separation is exactly what the company ties to the term "built-in security": protection is embedded not on top of the infrastructure as a separate product that can be installed or skipped, but inside it, at the hardware-software level of network equipment through which all traffic between agents, applications, and data stores inevitably flows.

Why do autonomous agents require a different approach to security?

Autonomous AI agents differ from conventional software in that they act with delegated authority: they run code, access external services, and make decisions with minimal human participation at each individual step. This creates a new risk profile — the threat may not come from an external attacker but from a standard, authorized tool that received too broad permissions or processed unvalidated data. The term "AI factories," which NVIDIA is promoting in conjunction with DOCA, describes data centers specifically designed for such workloads — with the expectation that the number of simultaneously running agents and the volume of internal network traffic they generate will grow orders of magnitude faster than in traditional corporate systems where the number of active processes is limited by the number of employees rather than the number of launched autonomous programs.

What this means for data center operators

For companies operating infrastructure under AI workloads, NVIDIA's emphasis on built-in rather than overlay security means they must rethink their security architecture: instead of relying solely on perimeter firewalls and access control at the individual application level, operators will need to build isolation and monitoring directly into the network equipment and DPU level through which traffic between agents and the rest of the infrastructure flows. The publication confirms a broader trend: as autonomous agents receive ever more rights to take real action — from code modifications to cloud resource management and orchestration of other agents — infrastructure producers like NVIDIA embed protection mechanisms not as an optional addition but as a mandatory component of the platform on which future "AI factories" are built, and thus competitive advantage in the AI infrastructure market as a whole.

In a broader context, the article stands alongside other NVIDIA initiatives promoting the term "AI factory" as the primary metaphor for describing next-generation data centers — the company has long positioned such facilities not merely as computational power but as industrial enterprises whose final product is intelligence expressed in tokens and actions of autonomous agents. Built-in DOCA security in this logic becomes not a secondary technical detail but part of the very definition of what it means to design a reliable "AI factory": without guarantees of isolation and traffic control between agents, no data center operator will be able to offer clients infrastructure suitable for industrial exploitation of autonomous systems.