China AI Platform Visits: Ask How Tenants Are Isolated
Multi-tenant AI needs evidence beyond separate interface colors. Examine document retrieval, session state, logs and administrative permissions in an authorized test.
Direct answer. Ask for an authorized two-tenant demonstration using synthetic data, then inspect the boundaries around retrieval, memory, logs and support access. A separate login screen does not by itself demonstrate isolation throughout the system.
Agree on the test authority
A China AI platform visit is not permission to probe the host's security. Obtain a defined demonstration scope and let authorized staff operate the test accounts. Use fictional organizations and documents. Never ask visitors to enumerate real customer records or bypass controls.
Create one synthetic fact available only in tenant A and ask tenant B a normal business question that would need that fact. Observe whether the platform returns an authorized limitation rather than an answer supported by the other tenant's material. One successful boundary test is useful evidence, not a penetration-test report.
Look beyond retrieval
| Boundary | Follow-up question |
|---|---|
| Documents | Where is tenant authorization enforced? |
| Conversation | Can memory or cached answers cross accounts? |
| Logs | Which team can inspect prompts and retrieved material? |
| Administration | How is emergency support access granted and recorded? |
Ask what changes when a user moves teams or leaves the customer organization. Cached permissions and previously created summaries may have a different lifecycle from the original documents. The presenter should identify what is demonstrated and what requires a separate audit.
Security and privacy evaluation themes in NIST AI RMF Core support these questions. The framework does not certify tenant isolation, and this walkthrough is not a security assessment.
Specify the evidence gap
Request a redacted architecture explanation, access-review process and independent assurance scope if available. Avoid treating a badge or a single blocked query as coverage of every subsystem.
The production-evidence guide helps label the boundary of your conclusion. Brief a platform visit with your security team before selecting demonstrations or requesting sensitive material.