liveBook · Manning

Tenant isolation

As explained in AI Governance

Tenant isolation is the separation of data, state, and operations belonging to different customers or organizational tenants in shared infrastructure.

A shared AI service can expose one customer’s information to another even when it does not retain data long term. Prompts and session state may briefly reside in shared memory, where an isolation failure can expose fragments of conversations or billing data. section 4.3.2

Tenant isolation separates the data and operations of customers, business units, or other user populations sharing infrastructure. For a retrieval system, the boundary must cover both where data is stored and what a request can retrieve. For centralized model serving, partition RAG indexes, prompts, secrets, and logs, and enforce tenant-scoped access at retrieval time. section 4.6.2 section 2.4.3

Combine isolation with controls that reflect individual permissions:

  • Partition data by tenant. A query runs against its tenant’s database partition rather than another tenant’s data. This creates a strong boundary but adds the complexity of managing multiple databases. section 4.6.2
  • Filter within a partition. Tag documents with security attributes such as department, sensitivity, region, or required clearance; filter results against the requesting user’s attributes before the AI sees them. section 4.6.2
  • Check current permissions at retrieval. An index can become stale when access is revoked or sharing links expire. Checking the source system or an authorization service reduces reliance on stale permissions, but adds latency and cannot protect data that should never have been indexed. section 4.6.2

Vendor assurances should be specific and testable. Ask how tenants are isolated in memory, at rest, and during inference; how long prompts remain in caches; and how incidents are disclosed and remediated. Also ask who can change system instructions and whether that access is audited. section 4.3.2 section 4.9

ClaimAssist: isolation before expansion

Before connecting an assistant to customer data, its team requested evidence and architectural details about the commercial vendor’s tenant isolation. Vague answers delayed a feature release. The assistant then queried records using the requesting customer’s identity, while fraud investigation files, personnel records, and financial account details were excluded from its collection to reduce exposure. section 4.9

How AI Governance covers it

How deep it goes
Explains partitioning, metadata filtering, and live authorization trade-offs, but leaves implementation details unspecified.
Examples
Uses a Redis session-state exposure and ClaimAssist’s vendor review and customer-scoped retrieval to illustrate isolation risks and controls.
What you can do after
Evaluate tenant-isolation controls for retrieval systems and vendor-hosted model infrastructure.
How it connects
Connects tenant isolation to retrieval-augmented generation, metadata filtering, access control, and real-time authorization.

Also explained in

Compare how 4 books explain tenant isolation →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of AI Governance. The section links point to the parts of the book it draws on.