Securing non-deterministic systems

I learned to code in BASIC on a Commodore 64 in the early 1980s. BASIC was actually a pretty neat high-level abstraction of assembly language, which itself was an abstraction sitting on top of machine code! Since then, I’ve seen a steady stream of higher-level abstractions - relational databases, object-oriented languages, package dependencies, open source libraries, cloud-native development, containers, serverless - that have made software developers more productive, and helped us build ever more powerful applications with them.

AI is the latest abstraction layer that sits on top of existing technology, making the underlying components more powerful and easier to use. A huge amount of R&D and startup capital in the past few years has gone towards exploiting this new abstraction.

The newest and best ideas naturally attract top talent and funding. Companies and startups building cool new AI-powered apps have their choice of tech stack. And many of them are creating their new software development tool chains around Cloudsmith, which is very satisfying to Cloudsmithers.

, who built an AI-powered platform pathologists use to diagnose cancers. They manage binary packages in Cloudsmith that they use to build for a wide variety of target environments. Or 

, who builds industrial computer vision applications that improve quality and productivity, using Cloudsmith to manage binary artifacts that go into their build pipelines. 

 provides a machine learning platform for document processing and automation, built on Cloudsmith for internal development and external distribution of containers and packages. And 

, who have built an AI-powered due diligence platform, using Cloudsmith as their single source of truth for binary artifacts.

Companies that build critical infrastructure for use by other AI companies are also gravitating to Cloudsmith. Take 

, who helps organizations use, manage, and govern AI metadata with their enterprise-grade context platform. They chose Cloudsmith to make sure their artifacts are reliably and securely distributed to enterprise developers. Or 

, who distributes their SDKs and runtimes via Cloudsmith for delivery of AI-driven perception solutions to robotics and computer vision tool companies. Then there’s 

, an open source AI search platform used to develop and run large-scale applications, combining big data, vector search, machine-learned ranking, and real-time inference to enable AI applications like retrieval-augmented generation, recommendation, and intelligent search at enterprise scale. Cloudsmith supports distribution for Vespa, streamlining delivery for their open source community as well as enterprise users.

And then it’s companies who deliver AI directly to the hands of end users. 

 calls itself “the world’s most private AI,” assisting users in real time across a wide range of personal tasks. They use Cloudsmith to build and distribute packages across the software supply chain.

Cloudsmith is best known for our enterprise-grade artifact management platform, which controls and secures the software supply chain for the world’s most demanding Fortune 500 and Global 2000 organizations. But there’s a reason we also show up so often as the platform of choice for innovative AI startups and spinoffs. These companies are typically run by lean teams of seasoned software engineers. They know what they need, and they want to do it right the first time. They want developer-friendly, well-documented tools that just work.

 to manage their Hugging Face models and datasets, and we’re increasingly powering our own development of new features with AI tools. But it’s also notable that the world’s most interesting AI companies are also choosing Cloudsmith - not just to manage ML models, but to power their entire software supply chain. That’s a vote of confidence that makes us proud to be Cloudsmithers.

Glenn Weinstein

The platform of choice for AI companies

Across your organization, teams are rapidly adopting AI and machine learning. They’re pulling ML models and datasets from public sources like Hugging Face and wiring them into workflows that are now reaching production. For platform and security leaders, this creates a familiar challenge: artifacts are entering the software supply chain outside established governance and controls.

Unmanaged AI artifacts carry real risks. They can’t be traced back to their source, they don’t follow compliance processes, and they often bypass the review gates you rely on for code and containers. Left unchecked, they become blind spots and liabilities in your supply chain. Without action, AI artifacts become the weakest link in your supply chain security.

The answer is to manage ML models and datasets with the same level of control and oversight as every other artifact in your supply chain. By managing them through a unified platform—with shared repositories, consistent policy enforcement, and full observability—you ensure that ML workflows are governed by the same standards as the rest of your software supply chain.

The growing adoption of ML models and datasets introduces both opportunity and risk. On one hand, models accelerate delivery by enabling new capabilities at unprecedented speed. On the other hand, they introduce opaque dependencies, security blind spots, and compliance risks into the supply chain—areas where governance practices are still catching up.

Some risks begin earlier in the lifecycle, like poisoned training data or tampered weights during model creation, which can compromise a model before it ever reaches a registry. But the most immediate challenge for platform and security leaders is governance once these artifacts enter the supply chain. Teams often pull models directly from public sources like Hugging Face, bypassing review and scanning. Without versioning and provenance tracking, it’s nearly impossible to prove which model is in use, where it came from, or what depends on it. And without consistent policies, organizations struggle to demonstrate compliance when auditors or regulators come calling.

These risks can’t be left unmanaged. The question is how to bring ML artifacts under control.

ML models and datasets are artifacts, and like any artifact, they need to be versioned, audited, and controlled with the same enterprise-grade policies you’ve put in place for packages like code libraries and container images. As organizations deliver production applications powered by AI, the priority for platform and security leaders is clear: maintain visibility, integrity, and trust in the components that make up these builds.

Cloudsmith’s new ML Model registry support

A Single Source of Truth for ML Models in DevOps

Cloudsmith extends artifact management to ML models and datasets, promoting them to first-class artifacts in your supply chain. This means you can control, secure, and distribute models alongside your language libraries, containers, and other binaries, all governed by the same policies and observability already in place.

: Require approvals, license checks, and compliance gates before a model or dataset is deployed.

: Every artifact is versioned, hashed, and immutable, so tampering can be detected and blocked.

: Capture where a model or dataset came from, how it has changed, and what depends on it.

: Manage ML models (PyTorch, TensorFlow, ONNX, and more) alongside other binary artifacts in the same repositories.

: Proxy and cache models and datasets directly from the Hugging Face Hub, while continuing to use the Hugging Face SDK for training and deployment — ensuring developers keep their workflows while you enforce governance.

DevOps has always been about balancing efficiency, stability, and security. The rise of ML workflows doesn’t change that balance - it heightens the need for it. By incorporating ML models and datasets into Cloudsmith, you extend the same supply chain controls that protect your code and containers to the AI artifacts now driving production systems.

The result: faster innovation without unmanaged risk, and a single platform to govern every component that ships to production.

Cloudsmith Blog

The platform of choice for AI companies

Extending supply chain governance to AI and ML artifacts