Going Beyond the Model: Why Open Source AI Matters
)
Summarize with AI
Inference. Expertise. Control.
Lifecycle solutions for your AI inferencing platform.
TAKE CONTROL OF INFERENCE
Mirantis joins The Linux Foundation’s Open Secure AI Alliance to help advance an open, inspectable and verifiable foundation for AI systems.
An AI system is never just a model. What is a large language model anyway?
At its simplest, it is a system that accepts input and produces a response. Early conversational systems demonstrated how powerful that interaction could be. But the transition from models that answer questions to systems that produce useful work came from what we build around them: a harness.
Agentic harnesses give models context, memory, tools and the ability to act. They connect models to source repositories, terminals, browsers, APIs, databases and business systems. They manage the loop in which a model observes, reasons, takes an action and evaluates the result.
The art and science of managing that interaction is becoming nearly as consequential as the models themselves.
An AI system is therefore no longer just a set of model weights. It is a complete stack: the model, its agentic harness, the tools and permissions available to it, the runtime in which it executes, and the infrastructure and software supply chain beneath it.
That is why Mirantis has joined the Open Secure AI Alliance, now hosted by the Linux Foundation. The alliance brings together organizations across open source, research, cybersecurity, cloud computing and enterprise software to develop and share open technologies.
Mirantis enters this effort with an opinionated perspective: the operational knowledge required to run AI safely should not remain in private runbooks, internal validation systems and the experience of individual contributors. We need to turn that knowledge into open artifacts that others can inspect, reproduce, validate and improve.
The AI system is the security boundary
Agentic systems do not merely generate text. They take action.
A coding agent might read and modify source code, invoke build systems, run commands and interact with deployment infrastructure. Other agents may retrieve sensitive information, call external services, access customer data or initiate business processes. The harness surrounding the model determines what actions are available, how tools are selected and what context is provided. That makes the harness a foundational part of the security model.
Because the harness shapes what an agent can do, it also shapes how an agent can be exploited. Prompt injection can manipulate an agent through untrusted content. Tool poisoning can cause it to invoke a capability under false assumptions. Excessive permissions can turn a mistaken action into a much larger incident. Credentials exposed inside an agent environment may be extracted or misused. Weak workload isolation can allow lateral movement between users, systems or tenants.
The risks continue below the harness too. AI stacks draw from models, containers, operators, plugins, drivers and libraries maintained across many different projects. A compromised artifact can enter through the software supply chain. Config drift can weaken an environment after it has been tested.
Securing the model alone cannot address these problems. The model, harness, runtime and infrastructure form a connected security boundary.
NVIDIA OpenShell illustrates the role of the runtime within that boundary. It runs autonomous agents inside isolated environments with explicit filesystem, process, network, credential and identity controls. The agent executes as a restricted process, while policy determines what it can access and where it can communicate. OpenShell does not replace secure agent design, identity management or infrastructure controls. It provides an open runtime boundary in which those controls can be enforced and inspected.
This layered approach matters. The harness determines what the agent is asked to do. Identity and policy determine what it is permitted to do. The runtime constrains what it can see, touch and execute. Infrastructure isolation limits the blast radius. Logs and evidence make its behavior reviewable afterward.
Ultimately, security does not reside in any of those layers. It emerges from how they work together.
Openness must reach the infrastructure layer
Below the agent runtime sits another complex system.
A production AI environment may include Kubernetes, GPU drivers, device plugins, network operators, container runtimes, storage systems, inference frameworks and observability services. Each component has its own lifecycle and compatibility requirements.
Small differences in kernels, drivers, operators or runtime configuration can cause failures that are difficult to diagnose and expensive to reproduce. Operators must accommodate changing generations of accelerated hardware, x86 and Arm architectures, bare-metal and virtualized deployments, multiple tenants and different workload requirements.
Historically, much of the knowledge required to make these combinations work has lived in internal validation pipelines, installation scripts and manual runbooks. Capturing it as open, versioned, verifiable artifacts lets operators bring accelerated infrastructure into production faster and keep it running reliably at scale.
To make that possible, open infrastructure should adhere to the following principles:
Open infrastructure makes its assumptions explicit
Configurations are declarative and reviewable
Versions and dependencies are recorded
Deployments are repeatable
Validations produce durable results
Operators can determine what was installed, where it came from, and whether the running environment still matches its intended state
In other words, operational experience must become reusable engineering artifacts.
Turning operational experience into open artifacts with AICR
At Mirantis, we needed to solve a practical problem: how to build, deploy and validate complete AI infrastructure stacks in a repeatable way for customers operating open-weight models.
NVIDIA AI Cluster Runtime, or AICR, provides an open foundation for that process.
AICR captures tested combinations of drivers, operators, kernels and system configuration as version-locked recipes. Those recipes can be rendered as deployment artifacts for Helm, Argo CD, Flux and Helmfile. After deployment, AICR can validate the running Kubernetes environment against the intended recipe.
AICR can also produce evidence that binds the validation result to the recipe, cluster snapshot, software bill of materials and test output. We developed an Apache 2.0 licensed Helm chart for NVIDIA AICR that turns this workflow into a standard, repeatable Kubernetes deployment. The chart consumes AICR without modifying it and makes the workflow deliverable through Helm, ArgoCD, Flux or a MultiClusterService for the k0rdent Kubernetes management platform.
We then contributed the integration to the upstream k0rdent catalog as a ServiceTemplate. This allows operators to provision Kubernetes clusters through k0rdent and use a MultiClusterService to install and validate the NVIDIA runtime across them. We tested the workflow end to end on an 8 GPU NVIDIA H200 bare metal system running k0s, a minimal, fully-contained Kubernetes distribution. It installed the selected AI stack, passed all four AICR deployment validation checks, and produced a portable evidence bundle that was independently verified.
This is how operational knowledge becomes open infrastructure. The runbook becomes a versioned recipe. The installation procedure becomes an automated deployment artifact. Environmental assumptions become explicit configuration. Test results become portable evidence.
More than a repository
A repository is an important starting point, but meaningful openness requires more than that. Users should be able to understand how a system is assembled. Operators should be able to reproduce a deployment in infrastructure they control. Contributors should be able to test improvements and report failures. Defenders should be able to inspect the evidence behind claims.
This advances the standard for open AI Infrastructure with proven GitOps and supply chain best practices:
Configuration should be declarative and reviewable
Dependencies and versions should be explicit
Deployment should be reproducible across environments
Validation should produce durable, inspectable results
Artifacts should carry provenance and integrity information
No single project or organization can deliver this end to end system. It requires collaboration among hardware and software AI companies and communities.
Why Mirantis is joining the effort
Mirantis has spent years working on the operational layers beneath distributed systems: Kubernetes, container runtimes, infrastructure automation, virtualization, bare metal, and lifecycle management.
Our work with accelerated infrastructure has reinforced how quickly complexity grows when these layers meet AI hardware and software. It has also made the limitations of closed operational knowledge increasingly clear.
Our participation in the Open Secure AI Alliance is therefore both practical and forward-looking.
We want to turn real operational experience into reusable open artifacts. We want infrastructure choices to remain available to users rather than being hidden behind closed appliances. And where our engineering work is not yet ready to support those goals, we intend to be transparent about what exists today and what still needs to be built. The AICR integration is an early example of that approach. As additional work matures, we intend to publish, document, and participate upstream.
Building the system around the model in the open
The next frontier of open AI is not limited to the model. It includes the harnesses that enable models to act, the tools and permissions that govern those actions, the runtimes that contain them, and the infrastructure and supply chains that support the complete system.
Each layer creates capabilities. Each layer also creates risk. Making those layers open allows engineers and defenders to inspect them, challenge the assumptions, reproduce the behavior and improve the overall system.
Mirantis has joined the Open Secure AI Alliance to help advance the work from the infrastructure upward. Our contribution will be grounded in what we know best: turning hard-won operational experience into open artifacts that can be deployed, tested and trusted by others.
)
)
)
)
