Hyundai tests Nvidia’s NemoClaw as Korean firms add agent security

Nvidia said in a post on X that Korea’s next wave of AI includes secure enterprise agents and expanded model access. According to the company, Hyundai Motor Group is exploring NemoClaw and OpenShell, and AIM Intelligence is integrating OpenShell.

The details come from an Nvidia blog post published around NVIDIA AI Day Seoul, which Nvidia said brought together more than 2,000 developers, startups and researchers.

Three Korean moves around agent safety and models

Nvidia says Hyundai Motor Group is exploring NVIDIA NemoClaw blueprints and the OpenShell runtime. The post gives no timeline, scope or use case for Hyundai’s work, and “exploring” is the only status Nvidia reports.

AIM Intelligence, a Korean AI security company, is integrating its Starfort agent control platform with OpenShell. Nvidia says the aim is to pair OpenShell’s runtime access controls with Starfort’s detection of malicious instructions and data leakage, to support agents used for employee productivity and enterprise workflows. The company is also considering a preliminary proof of concept for the public energy sector, with possible later uses in enterprise, industrial and robotics settings.

On the model side, LG AI Research is opening API access to K-EXAONE-2.5, which Nvidia says was built using Nemotron datasets, including Nemotron-Personas-Korea. That dataset holds 7 million fully synthetic Korean personas grounded in official Korean statistics, and is meant to anchor training data and evaluation in Korean contexts. Nvidia’s blog does not say when the API opens, who gets access or what it costs.

Nvidia describes OpenShell as a part of its Open Agent Safety Platform, offering sandboxed execution and policy-based controls to protect sensitive data, manage access to enterprise tools and govern what autonomous agents do.

What the sandbox actually does, and how early it is

According to Nvidia’s developer documentation and blog, OpenShell is Apache 2.0 based and sits between an agent and the infrastructure. Instead of relying on prompts to keep an agent in line, it enforces limits on the environment itself, with a policy engine covering the filesystem, network and process layers. Nvidia’s developer blog says agents such as Claude Code, OpenAI’s Codex and OpenClaw can run unmodified inside it.

NemoClaw is a layer on top. Nvidia’s docs call it an open-source reference stack for running always-on agents more safely inside OpenShell containers. The same documentation describes it as an early-preview stack meant for a trusted operator on one host, and earlier versions warned that APIs and configuration schemas may change between releases. That matters for readers weighing the Hyundai news: a large manufacturer exploring a stack that its maker still labels early is an evaluation, not a production rollout.

Who it affects in practice

For Korean enterprises, the practical change is a security layer that can be tested now and that local vendors are building around. AIM Intelligence’s integration is the clearest sign: a domestic security product will plug into the runtime rather than compete with it. A public energy sector proof of concept would test whether the combination holds up in a regulated environment, but it remains only a possibility in Nvidia’s wording.

For developers, the open license and the claim that popular coding agents run unmodified mean OpenShell can be tried without rewriting an agent. Teams building Korean-language products also get a path to LG’s K-EXAONE-2.5 through an API, plus a synthetic persona dataset they can use to test models against Korean contexts. None of Nvidia’s claims about the combined security benefits have been independently verified, and the post shares no test results, detection rates or latency figures.

The announcements also fit a wider Korean push that Nvidia is showing at the event. Earlier on the same day it said new Dynamo deployment recipes support Korean open models, and we reported that Korean chip startup Rebellions will connect its R200 NPUs to Nvidia’s platform through NVLink Fusion, with systems expected in 2028. Agent security is the software side of that same bet: Nvidia wants Korean hardware, models and governance tooling to sit on its stack.

Details the post leaves out

Nvidia does not give pricing, availability dates for the K-EXAONE-2.5 API, deployment timelines for Hyundai or AIM Intelligence, or any statement from the companies involved. It is also unclear what Hyundai would use agents for, whether NemoClaw or OpenShell alone is under evaluation for specific teams, or how OpenShell’s controls would be audited. Until those details arrive, the news is best read as a sign of interest from Korean enterprises rather than a measured result.