AI Enable Kit
Developers
LLM09 · OWASP LLM Top 10 2026

Vector and Embedding Weaknesses

Not applicable by designMoved from #8 to #9 in 2026

Security gaps specific to retrieval systems — reconstructing source text from embeddings, cross-tenant leakage in a shared vector store, or poisoned embeddings.

What it means

These are weaknesses unique to RAG (retrieval-augmented generation). Embeddings can sometimes be inverted to reconstruct the original text; a vector store shared across tenants can leak one customer’s data into another’s results; and poisoned embeddings can steer retrieval toward attacker-chosen content.

Why it matters

As RAG becomes a default pattern, a mis-scoped or shared vector database is an easy way for sensitive data to cross a boundary it never should.

How we approach it

Not applicable to this architecture. There is no vector database and no embedding store — the assistant does not retrieve from an indexed corpus. It works with your live data through your APIs at request time, which removes this entire class of risk.

If you add your own RAG layer inside your application, these controls — tenant isolation of the index, access control on retrieval, and guarding against embedding inversion — become your responsibility.

Frequently asked

Do you use embeddings or a vector database?

No. There is no vector store or embedding index. The assistant reasons over your live data through your APIs, so the vector-and-embedding risks do not apply here.

What if I need RAG in my application?

You can build it on your side. Just note that the retrieval index, its tenant isolation, and its access controls then become yours to secure.

← LLM08Hidden Context ExposureLLM10 →Improper Output Handling

All ten risks