Is the data an AI system can read (retrieval stores, training data, tool outputs) limited to what its users may see?
Facts read 16 Nov 2026. Page reviewed 16 Nov 2026.
Question 26 of the assessment asks: Is the data an AI system can read (retrieval stores, training data, tool outputs) limited to what its users may see? A retrieval system that reads with its own broad access can show a user data the user was never allowed to open.
Start the assessment ยท All 40 questions
Why it matters
A retrieval system that reads with its own broad access can show a user data the user was never allowed to open.
Answers and what each is worth
| Answer | Value |
|---|---|
| Yes, access follows the user | 1 |
| Scoped to the data store, not the user | 0.5 |
| No or not known | 0 |
How to check it in each cloud
- AWS
- Check knowledge base data sources, vector stores and S3 buckets used by Bedrock for the roles that read them, and whether metadata filtering enforces user permissions.
- Azure
- Check AI Search indexes and data sources for security trimming by user or group, and which identity reads the source storage.
- Google Cloud
- Check Vertex AI Search and RAG corpora for data store access controls, and which service account reads the underlying buckets.
The fix
Make retrieval and tool calls check the end user's permissions, and split sensitive data into separate stores. Effort: high.
Where it comes from
- Google Cloud security best practices AI and ML security: Google Cloud AI and ML perspective: security
- Microsoft cloud security benchmark AI-4: source
Related pages
Sources
- Google Cloud security best practices AI and ML security, read 16 Nov 2026
- Microsoft cloud security benchmark AI-4, read 16 Nov 2026
Last reviewed 16 Nov 2026.