Skip to main content

Restrict Retrieval of Raw Embedding Values

CCC.Vector.CN08 · Data Resilience

Ensure that stored vector values are not returned to query clients, so that embeddings cannot be harvested and inverted to reconstruct source data.

Related Capabilities

IDTitleDescription
CCC.Vector.CP01Embedding StorageThe service always stores high-dimensional vector embeddings, derived from raw input such as text, images, or audio, together with the record they represent.
CCC.Vector.CP03Similarity SearchThe service always answers nearest-neighbor queries by taking a query embedding and returning the stored vectors most similar to it.

Related Threats

IDTitleDescription
CCC.Vector.TH01Source Data is Reconstructed from Stored EmbeddingsEmbeddings stored by the service or returned by similarity queries may be probed repeatedly by a user with query access, or exported in bulk, and inverted with a reconstruction model. The text, images, or personal data from which the embeddings were derived are recovered, and the presence of specific records in the index is inferred. Confidentiality of the source data is lost even though the raw data was never stored in the service.

Assessment Requirements

IDTextApplicability
CCC.Vector.CN08.AR01When a similarity query is answered, the service MUST NOT return raw vector values unless the index is explicitly configured to expose them.tlp-amber, tlp-red

Guideline Mappings

FrameworkIDRemarks
FINOS-AIGFAIR-DET-001AI Data Leakage Prevention and Detection