Nvidia Expands DPU Role for Agentic AI Control

Nvidia is introducing a new infrastructure category called scale-in, designed to extend security controls across AI factories as agentic workloads grow more complex. Unlike earlier AI systems that simply answer queries, agentic systems make repeated model calls, retrieve data, and take actions, requiring tighter enforcement of boundaries. Nvidia is expanding the role of its BlueField data processing unit and DOCA software, alongside a new open-source runtime called OpenShell, which enforces agent permissions outside the workload itself. OpenShell supports sandboxed execution, credential management, and policy analysis across frameworks including Codex, Claude Code, and Hermes.