Hudson Buzby says AI services are reaching core production systems and need the same development, security, governance, and legal scrutiny as other software.
2
Open source models create risks through licenses, malicious packages, compromised model code, and unclear organizational usage, so enterprises need curated catalogs and central records.
3
A centralized AI gateway can control model access, route requests to suitable models, apply policies, and make model usage more portable across clouds and on-premises systems.
Summary
Hudson Buzby describes the gap between fast-moving AI development and the controls expected in enterprise software. Data science teams often worked outside standard DevOps practices, but AI services are now becoming core parts of products. He argues that these services need shared pipelines, testing, monitoring, security checks, governance, and legal review. Open source models make the problem harder because enterprises must assess model licenses, fine-tunings, packages, and possible malicious code. Buzby recommends a curated catalog, a proxy or gateway, and an ML bill of materials that records model and environment metadata. He also explains why organizations need visibility into every place generative AI is used, including codebases and third-party SaaS tools. The conversation ends with the idea that generative AI still fits many established MLOps practices, while a gateway can manage access, cost, model changes, and security policies across cloud and on-premises deployments.
Generative AI has shortened the path from experimentation to production
Hudson Buzby says traditional machine learning environments are often stitched together and barely holding on, while generative AI has made almost everyone consider their systems to be in production. Teams can launch agentic services like serverless functions and postpone testing and failure handling until something breaks. He expects development patterns to become closer to familiar software development life cycles as teams experience failures and standards emerge. AI assistants can write tests, but people still have to run the tests, inspect the results, and fix the problems.
AI applications need to move under the same controls as other software
Buzby argues that data science teams cannot continue with separate pipelines and looser standards once their AI services become fundamental products. Keeping separate processes does not make functional sense, and it creates security, governance, and legal risks. He expects AI applications to receive the same scrutiny as other production software. That includes shared development practices and stronger controls around services that previously looked like small experiments or weekend projects.
Enterprises want open source models, but they cannot allow unrestricted downloads
Buzby says large organizations have widely adopted managed providers such as OpenAI and Anthropic, while also looking toward open source models because of cost, privacy, competition, and a desire for more control. Their developers may want to use new models immediately, but enterprises need to examine licenses and governance first. Fine-tuned models on Hugging Face can carry different licenses from their source models, leaving legal teams unsure which terms apply. These practical questions can stop development even when the model itself is useful.
A model can carry ordinary software vulnerabilities
Buzby explains that a compromised Hugging Face model may contain more than poisoned data or hidden behavior. Its package, model code, or dependencies can include an exploit, a reverse shell, or code that sends credentials out of the environment. He compares this with scanning Docker images, Python packages, Conda environments, Poetry dependencies, and other software artifacts. Organizations need to scan both the model and the packages used to deliver it because either path can introduce malicious code.
Model access should pass through a curated catalog and gateway
Buzby recommends shrinking the set of models developers can access directly, rather than allowing every team to pull from Hugging Face or Ollama. A proxy can expose a reviewed and approved collection, record the model cards and safe tensor files used, and provide an audit trail. An ML bill of materials may not offer much immediate value, but it can help investigate a security incident or answer questions when licenses become enforceable later. A central registry is easier to audit than a search through notebooks.
Organizations need visibility into every use of generative AI
Buzby says companies often do not know where generative AI appears in their applications. JFrog can inspect an organization's code and identify packages, libraries, and SDKs associated with generative AI, which helps with licensing, vulnerability response, and future regulation. Demetrios Brinkmann points out that this misses marketing and other SaaS tools that use AI behind the scenes. Buzby agrees that those tools create another classification problem. He expects almost every service to include some form of generative development within three to five years.
MLOps practices still apply to language models and small models
Buzby says managing small language models remains MLOps, even when teams spend less time on experimentation and more time building deployable artifacts. Organizations still need A/B testing, persistent result data, queryable records, and drift analysis. Healthcare and other regulated customers may need every generative AI response stored. Mature MLOps systems give organizations a better base because they already support repeatable deployment and a system of record.
A central gateway can manage security, cost, and model portability
Buzby describes a gateway as a control point for model requests. It can decide which users may access particular models, route simple requests to smaller and cheaper models, apply security policies, and make it easier to replace one model with another. This matters because models will continue to change in cost, latency, size, and capability. JFrog's AI catalog is intended to make approved models available while allowing administrators to block or redeploy services that use models an organization has not approved.
"At the core, all of these organizations are looking for a platform, a standard that will allow them to easily define workflows in a safe, reliable, repeatable way."Hudson Buzby40:07
Who should watch
You are responsible for approving models, packages, or AI services in an enterprise and need a practical governance approach.
Your developers use Hugging Face, Ollama, hosted model providers, or coding assistants without a central record of what is running.
You are extending an existing MLOps platform to support language models across cloud, on-premises, or air-gapped environments.
