Advancing AI Infrastructure for Agentic AI with NVIDIA DOCA In-Silicon Security

The burgeoning era of AI, particularly with the rise of autonomous AI agents, necessitates a fundamentally new approach to infrastructure security. Traditional security architectures are ill-equipped to handle the scale, complexity, and performance demands of what NVIDIA terms 'AI factories'—environments that transform data into intelligence at unprecedented rates. These AI factories present a novel attack surface, encompassing infrastructure, software supply chains, models, data, and the agents themselves, which increasingly possess the authority to act autonomously. NVIDIA's response to this challenge is a robust in-silicon security framework, integrating NVIDIA BlueField Data Processing Units (DPUs) with the NVIDIA DOCA software stack. BlueField DPUs are purpose-built for AI infrastructure, combining high-performance networking, programmable compute, hardware acceleration, and advanced security capabilities into a single platform embedded within each AI factory compute node. This creates a hardware-enforced, workload-independent security layer that operates within its own trusted execution domain, isolating infrastructure and security services from potential compromises on the host system. The DOCA security stack provides a unified framework for protecting the entire AI factory. Key components include DOCA Argus, which establishes behavioral profiles for AI workloads to enable real-time monitoring for deviations indicative of compromise or malicious activity. DOCA Vault embeds storage security directly into the infrastructure layer, enforcing policy-driven data access even if the host operating system or applications are compromised, crucial for protecting sensitive datasets and AI models in multi-agent environments. Furthermore, DOCA Flow enables accelerated network enforcement, including Layer 7 firewalls and AI security gateways, to prevent threats from propagating laterally across systems while maintaining the high-speed performance required for large-scale AI training and inference. This comprehensive, full-stack security operates at AI speed and scale, providing end-to-end protection without sacrificing performance or scalability.