Enterprise AI agents are becoming more useful precisely because they can reach business tools, data and workflows that ordinary chatbots cannot. That access also turns runtime controls into an operational requirement. NVIDIA (NASDAQ: NVDA) said its OpenShell runtime is being combined with AI-security controls in Korea to support enterprise agents with sandboxed execution, policy-based access management and protections against malicious instructions and data leakage.
The work centers on NVIDIA OpenShell, which the company positions as a secure runtime within its Open Agent Safety Platform. NVIDIA said the software provides sandboxed execution and policy controls designed to protect sensitive data, manage access to enterprise tools and govern autonomous agent actions. The integration shows the type of infrastructure enterprises will need as agents move from advisory roles to carrying out multistep work.
AIM Intelligence, a Korean AI-security company, is integrating its Starfort agent-control platform with OpenShell. NVIDIA said the objective is to combine OpenShell’s runtime access controls with Starfort’s detection of malicious instructions and data leakage. The companies are considering an initial proof of concept for the public-energy sector, with potential applications in enterprise, industrial and robotics settings.
For IT leaders, the relevant question is not whether an agent can reach a system, but under what conditions it may do so. An agent connected to internal applications can retrieve sensitive information, make changes or pass data to other services. Sandboxing can isolate where code runs, while policy controls can define the tools and permissions available to that agent. Detection for harmful prompts and unexpected data movement adds another layer, although it does not eliminate the need for identity controls and human oversight.
NVIDIA also said Hyundai Motor Group is exploring NemoClaw blueprints and OpenShell for secure AI agents used in employee productivity and enterprise workflows. The company did not describe a production deployment or specific operational results. Enterprises evaluating similar designs should distinguish exploratory integrations from generally available, proven controls and test how policy enforcement behaves across their own applications and access models.
The announcement also links the runtime discussion to model development. LG AI Research is opening application-programming-interface access to K-EXAONE-2.5, a model developed using NVIDIA Nemotron datasets. NVIDIA described its Nemotron offering as open models, datasets and techniques for building specialized AI agents. Model selection can affect an agent’s behavior, but the OpenShell approach concentrates on a separate problem: constraining how that behavior interacts with enterprise systems at runtime.
That separation is important because companies increasingly need governance that follows an agent into execution, rather than only screening a model before deployment. Security teams will still need to establish approval paths, retain audit evidence, define acceptable actions and respond when controls detect a problem. The integration does not replace those processes; it offers components that can support them.
Compared with legacy AI deployments that focused on model access and content filtering, agentic systems require organizations to control an active software process with tools and permissions. NVIDIA’s OpenShell integrations point to a market where runtime boundaries, data-flow checks and scoped access become central differentiators for enterprise AI platforms—not optional add-ons after an agent is connected to production systems.