vector and embedding weaknesses
The OWASP risk area covering flaws in how vectors are stored and retrieved, such as one customer's data leaking into another's results or poisoned documents in an index.
A shared index without permission filters can hand one user another user's private documents.
In more detail
Filter retrieval by the requesting user's permissions inside the search, not afterwards in the prompt, and control who can add documents to an index. It is ninth in the OWASP Top 10 for LLM Applications 2026.
Related terms
- metadata filtering — Restricting a search to items whose metadata match conditions, such as product, language or access rights.
- vector database — A database or extension that stores embedding vectors with their metadata and answers nearest-neighbour queries.
- sensitive information disclosure — An application revealing personal data, secrets or confidential business information through a model's output.
Taught in
- Threat-Model An LLM App — AI Engineering: LLM Apps, RAG & Agents
References
- OWASP GenAI Security Project: Top 10 for LLM Applications 2026
- Qdrant documentation: Indexing (HNSW and payload indexes)