Skip to content
StrikeCyberStrikeCyber
Field Manual · Attack technique

Vector and Embedding Weaknesses

Reviewed November 13, 2026Attack technique 2 min read

Quick answer

Vector and embedding weaknesses are security flaws in the retrieval layer of AI applications: the vector database, the embeddings stored in it and the pipeline that fills it. They let users retrieve documents they should not see, let attackers plant content the model will trust, and can expose source text through the embeddings themselves. They are ranked eighth in the OWASP Top 10 for LLM Applications (2025).

On this page
01

How It Works

Retrieval-augmented generation (RAG) converts documents into embeddings, numeric representations of their meaning, and stores them in a vector database. When a user asks a question, the most similar chunks are retrieved and handed to the model. Several things go wrong. Access control is often lost in translation: documents from many sources and permission levels land in one index, and filtering by user or tenant is done with metadata that the application may not enforce correctly. Permissions are copied at ingestion and drift as the source system changes. Research has shown that embeddings can be partly inverted to recover the original text, so the vector store is as sensitive as the documents. And anyone who can add content to an indexed source can plant instructions or false information, which is RAG poisoning.

02

Why It Matters

Vector databases are newer than most enterprise data stores and are frequently deployed with weaker controls: shared indexes for many customers, broad service credentials and sometimes no authentication on internal or cloud-hosted instances. Because the retrieval layer decides what the model sees, a weakness here becomes sensitive information disclosure or manipulated answers for every user of the assistant.

03

How to Test for It

  1. Check how tenants and permission levels are separated: separate indexes or namespaces, or a shared index with metadata filters, and whether the filters are enforced server-side.
  2. Plant marker documents at different permission levels and test retrieval from users who should not see them.
  3. Revoke a user's access to a source document and test whether the assistant still retrieves it.
  4. Test whether the vector database is reachable directly, and with what credentials, from the internet or internal networks.
  5. Add documents containing hidden text, instructions or conflicting facts to an indexed source and see whether they reach answers.
  6. Review whether embeddings, backups and exports are protected to the same standard as the source data.
04

How to Reduce the Risk

Make retrieval permission-aware using the requesting user's identity, and prefer separate indexes or namespaces per tenant over shared indexes with filters. Re-sync permissions from source systems frequently and remove documents promptly when access changes. Require authentication and network restrictions on vector databases, encrypt and back them up as sensitive data, validate and sanitize content at ingestion, record which sources each answer drew on, and restrict who can add content to indexed locations.

05

Regulatory and Framework Context

In the US, joint CISA guidance on AI data security covers protecting data across the AI lifecycle, including retrieval stores, and the NIST AI RMF Generative AI Profile (NIST AI 600-1) treats data privacy and information integrity as core generative AI risks.

Common questions

Are embeddings safe to store because they are just numbers?

No. Research has shown that meaningful text can be recovered from embeddings. Treat the vector store as sensitive as the documents it was built from.

What is the most common RAG security flaw?

Retrieval that ignores the user's permissions, so the assistant answers from documents the person could never open directly.

Change log

  1. November 13, 2026First published.
(877) 657-8496