Open-weight AI: what new security guidance means for your business

What has changed?

Microsoft and NVIDIA have released a joint whitepaper, Risk Management for Open-Weight Models in Enterprise. It is the first in a planned series of resources to help organisations assess AI risks and choose suitable safeguards.

Open-weight models give organisations access to the underlying model weights: the numerical settings that shape how a model responds. That access offers more flexibility to inspect, test, adapt and run models in their own environments. However, greater control can also bring greater responsibility for security.

The paper introduces a shared way to categorise risks, drawing on established ISO standards and the NIST AI Risk Management Framework. Its central message is that access to a model’s weights does not, by itself, tell you whether an AI system is safe.

Why this matters to smaller businesses

For a small business in London or Manchester, the practical question is not simply whether an AI model is open or closed. It is what the system will do, what information it can access and what could happen if it behaves unexpectedly.

An internal drafting assistant needs a different assessment from a customer-facing system. Connecting either to business information or external tools introduces further considerations. Testing the model alone cannot establish that the finished application is safe.

Open-weight does not necessarily mean running everything yourself: these models can also be accessed through managed platforms. Responsibilities depend on the service and its configuration. If your business operates its own deployment, it may need to provide safeguards that a managed service would otherwise supply.

A clearer way to discuss risk

The guidance groups risks under five headings: cybersecurity, data protection, content, human–AI interaction and bias. This gives technical staff, business leaders and other advisers a common starting point for discussing concerns.

These categories are not a ready-made approval checklist. Businesses need to select the risks relevant to their intended use, decide how to test them, identify suitable controls and assign someone responsibility for each.

What to do next

Start by documenting the proposed use. Record where the model comes from, who will use it, whether customers will interact with it, and which data and tools it can reach. Include the permissions and operating environment of the wider application, not just the model.

Next, ask your IT team or provider to map relevant risks to tests and safeguards. Depending on the risk, measures could include access restrictions, filtering inputs and responses, monitoring or human review. Confirm which controls your provider supplies and which remain your responsibility.

Use test results as evidence, rather than a blanket guarantee. Microsoft Foundry offers built-in and custom evaluations, automated adversarial testing and configurable content-safety controls. If you use it, check that the capabilities you need are available for your chosen model and deployment, including whether they are still in preview.

Finally, agree when assessments will be revisited. Changes to the model, its connections or its purpose can affect the risks and the safeguards required. Testing should sit alongside ongoing monitoring and human oversight.

Useful guidance, with limits

This first whitepaper concentrates mainly on risks within the model itself. Microsoft and NVIDIA plan to expand their guidance to the surrounding software and operating layers, with further evaluation resources shared through a joint repository. Use the paper as a starting point, not as a substitute for assessing your complete business application.