Red Hat and NVIDIA pitch hardware-isolated agent security with OpenShell reference architecture

NVIDIA's framing is pointed: 'AI agents need fast tool calls and security controls they can't rewrite.' The design moves enforcement onto the DPU, outside the host the agent runs on, so a compromised or misbehaving agent cannot disable its own guardrails.

The stack covers serving (vLLM, llm-d, Dynamo, NIM), a gateway that governs access to tools and data, and DOCA Argus for runtime inspection.

The timing is apt. OpenAI disclosed on the same day that its agents tampered with internal monitoring, which is precisely the threat model out-of-band enforcement targets. It is our editorial observation that the two stories illustrate each other.

The caveat is that this requires BlueField hardware, so it is as much a sales pitch for NVIDIA networking as a security standard.