AIM Intelligence shows enterprise AI agent control with NVIDIA at AI Day Seoul

AIM Intelligence presented a joint session with NVIDIA at AI Day Seoul 2026 on November 9, showing how enterprises can see, govern and verify employee AI agents. The demo linked AIM Intelligence’s Starfort AIDR and Stinger tools to NVIDIA OpenShell, with plans to extend the approach to public-sector, industrial and robotics use cases.

Why it matters: - Enterprises are handing real work to AI agents across code, documents, CRM systems and email. - That shift creates a visibility gap around permissions, behavior and data handling. - AIM Intelligence says its control stack is designed to help organizations reduce data leakage, indirect prompt injection and unsafe agent actions.

What happened: - AIM Intelligence announced a joint session with NVIDIA at NVIDIA AI Day Seoul 2026, held Nov. 9 at COEX in Seoul. - The session, titled "Agentic AI in Production: From Open Models to Agents You Own," showed a configuration that connects AIM Intelligence technology with NVIDIA OpenShell, NVIDIA’s security runtime for AI agents. - The demo was presented by Hanwool Lee, Head of Defensive Security at AIM Intelligence, and Jinho Lee, Solutions Architect at NVIDIA. - AIM Intelligence said the setup lets enterprises see the AI agents employees use, govern them under organizational policy and continuously verify their safety.

The details: - OpenShell runs agents in an isolated execution environment and defines which files, network destinations and processes the agents can reach. - Starfort AIDR inspects content moving inside that boundary and checks whether it fits the user, data and business context and the organization’s policies. - A request to the same approved service can be allowed when it carries a public summary and blocked when it carries the original customer list. - Stinger repeatedly runs attack scenarios to verify whether the controls actually work. - Stinger places the attack executor and the test target in separate OpenShell sandboxes so blocked attacks can be separated from genuine defensive success. - Stinger also re-tests under the same conditions whenever models, tools or permissions change. - In roughly 1,920 internal Stinger tests under the stated test conditions, out-of-scope executions observed 24 times before OpenShell policies were applied were observed zero times after. - AIM Intelligence said that figure reflects observations within that test scope and policy set and does not represent a general block rate. - The presentation also outlined plans to connect NVIDIA SkillSpector and NVIDIA NeMo Guardrails to the configuration. - AIM Intelligence said the OpenShell and Starfort AIDR integration is currently a proposed configuration and is not a completed integration or jointly released product.

Between the lines: - The demo points to a broader enterprise shift from deploying standalone agents to managing them as governed systems. - The emphasis on separate sandboxes for testing suggests that validation is becoming as important as blocking behavior in production agent security. - The approach also reflects a move toward policy enforcement at the execution layer, not just at the model or prompt layer.

What's next: - AIM Intelligence is reviewing plans to extend the approach to public-sector and industrial agents. - The company is also looking at physical AI environments such as robotics. - In the public energy sector, AIM Intelligence is reviewing a plan to work with KEPCO KDN on pre-validating a cyber threat response framework for an energy-specialized AI data center. - That work would use a simulated environment and de-identified samples rather than live operational data. - The planned testing would focus on instruction contamination, approval bypass and sensitive data leakage risks in automated operations. - AIM Intelligence also said it presented a case at a Samsung Group affiliate where agents were identified through Shadow AI discovery and brought under enterprise control through Starfort AIDR.

The bottom line: - AIM Intelligence is positioning AI agent control as an enterprise security layer that can inspect, constrain and continuously test agent behavior before organizations scale it further.