NVIDIA described two Korean moves around agentic AI in a post on its blog: one on agent security, one on a model.
Agent security: OpenShell and Starfort
OpenShell is a component of the NVIDIA Open Agent Safety Platform. It is a runtime that executes agents in a sandbox and applies policy-based controls. Those controls protect sensitive data, limit which enterprise tools an agent can reach, and govern the actions an autonomous agent takes.
AIM Intelligence, a security firm specializing in artificial intelligence based in Korea, is merging its Starfort platform for managing agents with OpenShell. The goal is to combine the runtime's access controls with Starfort's identification of harmful instructions and data exfiltration. The firm is additionally weighing an initial proof of concept aimed at the public energy sector. According to the source, the approach could later apply to enterprise, industrial and robotics settings. That is a possibility, not a commitment.
Hyundai Motor Group is exploring NVIDIA's NemoClaw blueprints together with the OpenShell runtime. The source gives no scope, timeline or results, so this is an evaluation and not a rollout.
Model side: K-EXAONE-2.5
LG AI Research is opening API access for K-EXAONE-2.5. The model was developed using NVIDIA Nemotron datasets, among them Nemotron-Personas-Korea. That dataset holds 7 million fully synthetic Korean personas built on official Korean statistics. Model builders use it to anchor training data and evaluation in Korean contexts.
The source does not give pricing, rate limits, benchmark results or a launch date for the API, so none of those can be assumed.
What is still open
The post does not say when the Starfort integration will be generally available, what it costs, or which OpenShell policies ship by default. Before building on either piece, ask the vendors for those details and test against your own data.