NVIDIA DOCA: Built-In Security for Agentic AI Infrastructure

NVIDIA has published a post on its corporate developer blog (NVIDIA Developer Blog) explaining how the NVIDIA DOCA platform provides built-in security for the infrastructure that runs autonomous AI agents. The post starts from the premise that the AI era is creating a new class of infrastructure: so-called "AI factories" that turn data into intelligence for autonomous agents operating at an unprecedented scale of tasks and level of independence.

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

• The platform is NVIDIA DOCA, a software framework for DPUs (data processing units)
• The target hardware is the NVIDIA BlueField family of cards
• NVIDIA's concept is "AI factories": data centers that produce intelligence as a product
• The target workload is autonomous AI agents with delegated authority
• The source is the NVIDIA Developer Blog

NVIDIA DOCA is the company's software framework for DPUs, BlueField-class devices that take over some of a data center's networking and infrastructure functions, offloading the main CPU and GPU compute resources. The idea is to move control over traffic, access and security policies to a separate layer isolated from the main compute workload, so that a compromised application or agent on one server does not automatically open access to the rest of the data center. NVIDIA ties this principle of separating layers to the term "built-in security": protection is not bolted on top of the infrastructure as a separate product that can be installed or skipped, but built into it, at the hardware and software level of the networking equipment that all traffic between agents, applications and data stores inevitably passes through.

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, call external services and make decisions with minimal human involvement at each individual step. This creates a new risk profile. The threat may come not from an external attacker but from a legitimate, authorized tool that was granted overly broad permissions or processed unverified data.

The term "AI factories," which NVIDIA promotes alongside DOCA, describes data centers designed specifically for such workloads, on the expectation that the number of agents running at once and the volume of internal network traffic they generate will grow by orders of magnitude faster than in traditional enterprise systems, where the number of active processes is limited by the number of employees rather than the number of autonomous programs running.

What this means for data center operators

For companies running infrastructure for AI workloads, NVIDIA's bet on built-in rather than add-on security means rethinking their protection architecture. Instead of relying only on perimeter firewalls and application-level access control, operators will need to build isolation and monitoring directly into the networking hardware and DPUs that carry traffic between agents and the rest of the infrastructure. The publication confirms a broader trend: as autonomous agents gain more rights to take real actions, from changing code to managing cloud resources and orchestrating other agents, infrastructure vendors like NVIDIA are building protections in as a mandatory component of the platform on which future "AI factories" will be built, rather than as an optional extra.

That, in turn, makes them a competitive advantage in the AI infrastructure market as a whole.

In a broader context, the post sits alongside other NVIDIA efforts to promote "AI factory" as the main metaphor for next-generation data centers. The company has long positioned such facilities not merely as computing capacity but as industrial plants whose end product is intelligence, expressed in tokens and in the actions of autonomous agents. In this logic, the security built into DOCA is not a minor 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 can offer customers infrastructure fit for industrial operation of autonomous systems.