As the AI landscape evolves, enterprises are increasingly choosing between open-weight and closed-weight models. While open-weight models offer significant flexibility for customization and inspection, they also shift the responsibility for risk management and safety safeguards onto the enterprise. Microsoft and NVIDIA have collaborated to address these challenges with the release of their new paper, ‘Risk Management for Open-Weight Models in Enterprise.’
Moving Beyond the ‘Open vs. Closed’ Debate
The central thesis of the new guidance is that the degree of openness does not inherently determine an AI system’s risk. Instead, risk is a combination of the specific model, the use case, the deployment environment, and the implemented safeguards.
A Standardized Risk Taxonomy
To help cross-functional teams—including engineering, security, and legal—speak the same language, the whitepaper introduces a shared taxonomy divided into five core categories:
- Cybersecurity
- Data Protection
- Content
- Human-AI Interaction
- Bias
This framework is designed to be adapted by enterprises to their unique operational needs, helping them map risks to concrete evidence and mitigation strategies.
Connecting Evaluations to Risk
Enterprises need evidence to manage AI risks effectively. The paper outlines how to use evaluations to quantify risks and validate controls. Tools like Microsoft Foundry play a critical role here, offering automated red teaming and content-safety controls that help detect harmful inputs and outputs. However, the authors emphasize that evaluating a model is only one piece of the puzzle; teams must also assess their data access, permissions, runtime environment, and operational controls.
Shared Responsibility
Because open-weight models allow for greater control over deployment, enterprises must clearly define who is responsible for maintaining safeguards. Whether using a managed platform or hosting models independently, organizations must ensure that controls are robust, monitored, and revisited as their applications evolve.
Future Outlook
This paper serves as the first in a series of joint resources from Microsoft and NVIDIA. Future work will expand into the runtime, harness, and application layers of AI deployments.
We encourage developers, security professionals, and researchers to read the whitepaper, explore the joint repository of evaluation resources, and engage with the growing community effort to standardize AI safety and risk management.
