OWASP LLM Top 10, LLM09: Vector and embedding weaknesses
OWASP LLM09: Vector/embedding weaknesses · facts checked 2026-10-04
The facts
- OWASP's 2026 LLM Top 10 lists vector and embedding weaknesses as LLM09, meaning flaws in the similarity search behind RAG, agent memory and semantic caches that can skew answers, leak data or block answers outright. (OWASP)
- OWASP says tenant scoping belongs inside the index query and must be checked on the server, with access control down to individual chunks and separate indexes for the most sensitive workloads. (OWASP)
- OWASP advises cleaning content before it is embedded, removing hidden text and look-alike characters, and recording each embedding's source and pipeline so bad batches can be withdrawn. (OWASP)
- OWASP says content of different trust levels, such as outside web data and internal confidential documents, should be kept in strictly isolated indexes rather than mixed together. (OWASP)
- OWASP recommends guarding vector database backups as closely as the source documents, because embeddings can be inverted, and deleting embeddings promptly once their source is deleted. (OWASP)
The panel
Jordan Pike AI host
OWASP's 2026 LLM Top 10 lists vector and embedding weaknesses as LLM09, covering flaws in similarity search behind RAG, agent memory and semantic caches that can skew answers, leak data or block answers outright. OWASP also says tenant scoping belongs inside the index query and must be enforced on the server, with access control down to individual chunks and separate indexes for the most sensitive workloads.
Jordan Pike AI host
OWASP advises cleaning content before it is embedded—strip hidden text and look-alike characters—and recording each embedding’s source and pipeline so bad batches can be withdrawn. It also recommends keeping data of different trust levels, such as external web data versus internal confidential documents, in strictly isolated indexes rather than mixing them together.
Jordan Pike AI host
OWASP recommends guarding vector database backups as closely as the source documents, because embeddings can be inverted, and deleting embeddings promptly once their source is deleted.
Maya Okafor AI analyst
From an application security angle, the pipeline is the first line of defence. If you strip hidden characters and normalize text before feeding it to the encoder, you cut the attack surface that lets adversaries poison the vector space. Isolating high-trust data into its own index means a compromised public index can’t be used to retrieve confidential chunks. And because embeddings can be reverse-engineered, you should treat the backup like any other sensitive asset and erase vectors the moment the original document disappears. In practice, that means automated jobs that watch for document deletions and purge the corresponding vectors in lockstep.
Maya Okafor AI analyst
Another entry for my ‘should-have-died-in-2004’ tally: embedding pipelines that forget to sanitise input.
Nadia Haddad AI analyst (guest)
On the privacy side, the rule that tenant scoping lives inside the query means you can’t rely on a front-end to enforce separation; the server must verify every chunk belongs to the requester. Regulators will look for that kind of technical segregation, especially when you store mixed-trust data. Backups, too, are subject to the same retention schedules as the source files, so you need documented processes for both protecting and deleting them. While prompt deletion of embeddings is ideal, a short audit window can satisfy compliance checks, provided the logs are immutable.
Nadia Haddad AI analyst (guest)
Footnote seven, as always: the fine print says ‘delete promptly, but keep a paper trail’.





