Microsoft and NVIDIA publish open-weight AI risk guidance

Introduction

Open-weight AI models give enterprises more flexibility to inspect, customize, and run models in their own environments. That flexibility can be valuable, but it also shifts how organizations must think about security, governance, and operational responsibility.

Microsoft and NVIDIA have now published Risk Management for Open-Weight Models in Enterprise, a joint paper designed to help organizations evaluate and mitigate risk based on the actual deployment context, not just whether a model is open- or closed-weight.

What’s new

The new guidance introduces a practical framework for enterprises deploying open-weight models:

• Shared risk taxonomy: Risks are grouped into five categories:
  • Cybersecurity
  • Data Protection
  • Content
  • Human-AI Interaction
  • Bias
• Deployment-focused risk analysis: The paper stresses that model risk depends on the use case, architecture, permissions, data access, runtime, and surrounding controls.
• Evaluation mapping: Microsoft and NVIDIA connect specific risks to evaluations that can provide evidence for risk decisions.
• Standards alignment: The approach is grounded in established frameworks, including ISO standards and the NIST AI Risk Management Framework.
• Joint repository: Microsoft and NVIDIA plan to expand the supporting resources with additional evaluations and technical guidance over time.

Why this matters for IT and security teams

The biggest takeaway is that weight openness alone does not define AI risk. An open-weight model used internally for drafting documents has very different risk implications than the same model embedded in a customer-facing workflow or connected to external tools and data sources.

For security and platform teams, this means model selection is only one part of governance. Teams also need to determine:

• Where safeguards should be implemented
• Which controls are provided by a managed service versus owned internally
• How to validate model behavior with testing and red teaming
• How to monitor changes as models, prompts, and deployment patterns evolve

Microsoft also highlights Foundry evaluations and content-safety controls as implementation options for assessing quality, groundedness, and safety, while helping detect harmful inputs, outputs, and prompt attacks.

Recommended next steps

Organizations evaluating or already running open-weight models should consider the following actions:

1. Review the Microsoft-NVIDIA paper and map its taxonomy to your internal AI governance process.
2. Assess deployments by context, including users, data access, connected tools, and business impact.
3. Identify control owners for model, runtime, application, and operational layers.
4. Use evaluations as evidence, not as a complete safety guarantee.
5. Plan for continuous review as models and deployment architectures change.

Bottom line

This guidance gives enterprises a more structured way to manage open-weight AI risk without reducing the conversation to open versus closed models. For administrators, security teams, and AI governance leaders, the paper offers a useful starting point for aligning risk assessments, evaluations, and safeguards across enterprise AI deployments.