This chapter addresses securing the embeddings and vector stores that act as semi-persistent and persistent "memory" for AI systems through Retrieval-Augmented Generation (RAG).
Fine-grained access controls and query-time scope enforcement must be applied to every vector collection.
| # | Description | Level |
|---|---|---|
| 8.1.1 | Verify that vector identifiers and namespaces enforce uniqueness per tenant and prevent cross-tenant collisions. | 1 |
| 8.1.2 | Verify that document metadata tags are immutable after the initial write. | 2 |
| 8.1.3 | Verify that retrieval operations enforce scope constraints. | 2 |
Content must be pre-screened before vectorization, and memory writes treated as untrusted input, to prevent ingestion of unsafe payloads.
| # | Description | Level |
|---|---|---|
| 8.2.1 | Verify that sensitive fields are detected before embedding and are masked, tokenized, or dropped. | 1 |
| 8.2.2 | Verify that vectors that fall outside normal clustering patterns are flagged and quarantined before entering production indices. | 2 |
| 8.2.3 | Verify that no component reachable from agent execution can write agent outputs or tool outputs to trusted agent memory without explicit source validation. | 2 |
| 8.2.4 | Verify that content crafted to manipulate retrieval results is detected and rejected or quarantined before vectorization. | 3 |
| 8.2.5 | Verify that new content written to memory is checked for contradictions with what is already stored and that conflicts trigger alerts. | 3 |
Retention and revocation must be explicit and enforceable for memory and RAG indices.
| # | Description | Level |
|---|---|---|
| 8.3.1 | Verify that expired vectors are excluded from retrieval results. | 2 |
| 8.3.2 | Verify that memory can be reset. | 2 |
| 8.3.3 | Verify that quarantined content is retained but excluded from all retrieval results. | 3 |