Hyundai and AIM Intelligence test Nvidia’s OpenShell sandboxes before letting agents near real systems

Agents that call tools and move data are only as trustworthy as the box they run in. Hyundai Motor Group is exploring Nvidia’s NemoClaw blueprints and the OpenShell runtime, part of Nvidia’s Open Agent Safety Platform (platform launch). Korean security firm AIM Intelligence is integrating its Starfort agent-control platform with OpenShell. On the model side, LG AI Research is opening API access to K-EXAONE-2.5, built using Nvidia’s Nemotron datasets including a Korean synthetic persona set. Together it is a small sign that secure deployment is becoming part of the agent pitch rather than an afterthought. These came alongside Nvidia’s Korean physical-AI news (physical AI).

OpenShell provides sandboxed execution and policy controls over data, tool access and agent actions (OpenShell). AIM Intelligence wants to layer detection of malicious instructions and data leakage on top. Runtime limits plus detection is the right shape: one stops an agent doing what it shouldn’t, the other notices when someone is trying to make it.

The caveats are real. Hyundai is exploring, not deploying. AIM Intelligence is considering a preliminary proof of concept for the public energy sector, an early stage in exactly the setting where a compromised agent could do physical harm. Sandboxes are only as good as their policies, and prompt-injection detection has a long history of being bypassed; Microsoft’s recent work on backdooring agents through a few neurons shows how deep attacks can go (Lazarus). Red-team results should come before anyone calls this secured.

Nemotron-Personas-Korea is an open set of 7 million fully synthetic Korean personas grounded in official statistics, a sensible way to make a model fluent in local context without scraping private data. The subtler risk is that synthetic personas flatten a population into its averages: whatever the statistics miss, the model may miss, and evaluations built on the same synthetic base can hide the blind spot. Independent benchmarks and developer uptake for K-EXAONE-2.5 will be the test. More in our Nvidia coverage and cybersecurity coverage.