1-minute read
What a defender needs to know
Microsoft and NVIDIA published guidance on October 16 for enterprises deploying open-weight models. Their five risk categories are Cybersecurity, Data Protection, Content, Human-AI Interaction and Bias. The work concentrates on model-layer risks and maps selected risks to evaluations; an agent also needs assessment of its permissions, data access, connected tools and operational controls. An internal drafting assistant and a customer-facing agent may use the same model with very different authority.
Inventory the model’s provenance, intended use, users and hosting arrangement. For every application using it, record the data stores and tools it can reach, the permissions it holds and the owners of the evaluations and safeguards. Operating the model yourself can put more safeguard work on your team; on a managed platform, check the actual service and deployment configuration. Restrict tool data access to the application’s job, and choose filtering, monitoring or human review against the risk in that workflow. Check whether any chosen platform control is available for your model and whether it is in preview.
A model-output evaluation does not test the application’s tool permissions. With a test identity and synthetic records, verify that the approved task succeeds, an out-of-scope tool call is denied, and monitoring records the attempt. Repeat the relevant tests after model, tool or permission changes. Record attempted tool use and changes to application access, then verify that monitoring captures both.