Control #

D

2

.

2

Run AI workloads only in approved regions

Restrict all model training, fine-tuning, and inference to infrastructure located in approved regions. This applies to production and non-production environments.

Evidence

Policy documentation defining model residency rules

Evidence of geo-restriction settings in cloud platforms

Cloud configuration logs demonstrate models are deployed in the correct region

Recommended actions

Model your geo-restriction strategy using the NIST information lifecycle

Use a framework like the NIST information lifecycle to guide your geo-restriction strategy:

1. Creation or Collection

  • Ensure data collection systems (e.g., user inputs, data ingestion pipelines) are hosted or routed through approved regions.

  • Log collection endpoints and verify alignment with data residency rules.


2. Processing

  • Validate that any preprocessing, transformation, or labeling is done in-region.

  • Ensure third-party services (e.g., enrichment APIs, annotation platforms) comply with geographic restrictions.


3. Storage

  • Use cloud storage policies to restrict datasets, model artifacts, logs, and embeddings to compliant regions.

  • Configure backup and disaster recovery storage to meet the same requirements.


4. Use (Inference / Training)

  • Confirm that both training jobs and inference endpoints are region-bound.

  • Cloud configurations (e.g., GCP, AWS, Azure) should enforce compute resource residency.


5. Dissemination

  • Monitor how outputs (e.g., predictions, summaries, generated content) are shared or served, especially across borders.


6. Destruction and Deletion

  • Ensure deletion policies apply at the region level, and verify that data sanitization and model decommissioning procedures are geographically scoped.

  • Include logs and temporary artifacts in data destruction workflows.



Provide feedback on this control