Insert content that embeds close to trusted documents or contains hidden instructions.
LLM08:2025
Vector and Embedding Weaknesses
Vector and embedding weaknesses affect RAG systems when embeddings, vector stores, chunking, metadata, retrieval filters, or similarity logic expose or distort context.
Step 01
Input
Step 02
Model
Step 03
Tool / Data
Step 04
Impact
What it is
Weak access controls or data-integrity checks in an embedding and retrieval pipeline can expose restricted material or place manipulated content in the model's context. Similarity to a query does not establish permission or source reliability.
Why it matters
Weak vector controls can create cross-tenant leakage, stale answers, poisoned context, hidden prompt injection, and decisions based on irrelevant or unauthorized material.
Failure path
One way this failure can develop.
Use this example to locate the input, permission or application check that would interrupt the sequence in your system.
Exploit weak metadata filters, tenant boundaries, chunking, or ranking behavior.
Cause unauthorized, poisoned, stale, or irrelevant chunks to steer the final answer.
Defenses
Controls worth checking.
Check where each control runs, who owns it and how you will test it again after a prompt, model or workflow change.
Enforce retrieval authorization
Apply tenant, role, document, and classification constraints before similarity search results reach the model.
Track chunk lineage
Carry source, owner, timestamp, trust level, and access metadata through ingestion, retrieval, answer generation, and citations.
Test adversarial retrieval
Probe similar-looking documents, poisoned text, stale chunks, hidden instructions, and cross-tenant collisions.
Signals and control gaps
- Answers citing chunks from the wrong tenant, role, source, or document class.
- High-confidence answers based on stale or low-trust material.
- Retrieved chunks containing instructions intended for the model.
Questions for your team
- Are vector filters mandatory and server-side constructed?
- Can users influence retrieved context through uploaded content?
- Does each answer preserve source and authorization lineage?