NVIDIA Targets AI Agent Security and Infrastructure Costs

AI agents can be expensive employees. They consume computing resources even when they’re waiting around, and giving them too much access can create a security headache.

NVIDIA wants to tackle both problems. At KubeCon + CloudNativeCon on Tuesday, the company announced plans to contribute its OpenShell security runtime to the Cloud Native Computing Foundation (CNCF), alongside new open-source tools designed to make AI infrastructure more efficient.

For businesses deploying autonomous agents, the announcements address two growing concerns: keeping AI systems under control and avoiding unnecessary infrastructure costs as deployments scale.

OpenShell gives organizations control over what autonomous AI agents can access and do. Its runtime enforces security policies as agents interact with tools, access inference services and execute workloads.

Moving the project into the CNCF Sandbox allows organizations building OpenShell into their products to participate in its development and governance. NVIDIA says hundreds of organizations have used the software, which has received tens of thousands of contributions in less than a year.

“With OpenShell governed by the CNCF, teams have a shared runtime to secure agents, which makes it the right place to solve data safety for agents once instead of tool by tool,” said Pranava Adduri, CTO and co-founder of Bedrock Data.

NVIDIA and Google target agent computing costs

NVIDIA is also contributing to Agent Substrate, a Google-originated CNCF Sandbox project designed to run long-lived AI agents on Kubernetes.

The framework addresses the computing resources consumed by agents waiting for instructions or model responses. It can suspend idle agent environments, preserve their state and resume them using shared computing resources when needed.

This could help organizations run more agent sessions without keeping dedicated infrastructure active for each one.

NVIDIA is working to integrate the framework with its Dynamo inference software and KAI Scheduler. NVIDIA and Google are also developing an integration between Agent Substrate and OpenShell to extend agent security across cloud and on-premises environments.

NVIDIA also introduced three open-source components for its DSX AI Factory Platform:

• NVIDIA Kubernetes Engine (NVKE): Automates Kubernetes cluster provisioning and management.
• NVIDIA Kubernetes Recipes (NVKR): Provides reproducible infrastructure components and version-pinned software images.
• NVIDIA Snapshot: Restores checkpointed AI workloads without repeating model loading and initialization.

NVKE and NVKR are available in public preview, with several infrastructure providers planning integrations. Snapshot has delivered notable results in partner testing. Crusoe reduced GLM-5.3 workload readiness time from roughly eight minutes to 31 seconds on NVIDIA HGX B300 systems.

Saturn Cloud also cut vLLM startup times from between two and 8.5 minutes to six to 34 seconds.

These results depend on compatible hardware and previously created checkpoints. They do not include full infrastructure provisioning.

What this means for enterprise AI

NVIDIA’s announcements address interconnected challenges facing companies moving AI agents into production: security, computing costs and service availability.

Keeping agents running continuously can waste expensive GPU resources, while insufficient access controls can expose enterprise systems to unwanted actions. Combining OpenShell’s governance capabilities with Agent Substrate’s workload management could help infrastructure providers address both problems.

However, the projects have different maturity levels. OpenShell’s governance transition is ongoing, Agent Substrate remains under development, and NVKE and NVKR are still in public preview.

How developers and AI users could benefit

Developers could gain more control over agent permissions and more efficient ways to manage long-running sessions. Enterprise customers could eventually see better GPU utilization and faster workload recovery.

For users of AI-powered business applications, faster workload recovery could mean fewer service interruptions. However, these announcements do not immediately change existing AI assistants or guarantee that every agent deployment is secure.

NVIDIA’s announcements address three challenges businesses face when moving AI agents into production: controlling access, managing infrastructure costs and keeping workloads available.

However, organizations should consider several limitations before adopting the technologies:

• Project maturity: OpenShell’s governance transition is ongoing, Agent Substrate remains under development, and NVKE and NVKR are in public preview.
• Compatibility: Snapshot’s reported performance improvements depend on supported infrastructure and previously created checkpoints.
• Security: OpenShell provides a framework for enforcing agent policies, but organizations must still configure appropriate permissions and monitor agent activity.

For enterprises evaluating these technologies, the next step is to test them against existing Kubernetes environments, security requirements and AI workloads before committing to production deployments.

NVIDIA’s broader ambition is to make enterprise AI infrastructure more open and efficient. Whether that translates into lower operating costs and more reliable applications will depend on how these tools perform beyond controlled demonstrations.

In other news, NVIDIA introduced GR00T 2, an AI model that learns robotic tasks from human videos and demonstrations, achieving a 95% average success rate with just five demonstrations per task in company testing and potentially reducing the training effort required for industrial automation.