LLM09: Vector and Embedding Weaknesses · 2026
Embedding geometry and similarity search enable cross-tenant leaks, inversion, poisoning, retrieval blocking or cache manipulation.
Summary
Embedding geometry and similarity search enable cross-tenant leaks, inversion, poisoning, retrieval blocking or cache manipulation.
How the attack works
Similarity search does not filter by tenant; it retrieves a document from another organization and includes it in the answer.
Example
Vulnerable · Pseudocodehits = vectors.search(embed(query))
reply = llm(query, context=hits)
# No tenant or document authorizationFix: Apply tenant scope and document authorization within the query; authenticate ingestion sources.
Prevention
- Apply tenant scope within the query.
- Authenticate sources and version every embedding.
- Do not expose raw scores and monitor anomalies.
Lineage
- 2025 · LLM08: Vector and Embedding Weaknesses Moves down to rank 9 in 2026
- 2026 · LLM09: Vector and Embedding Weaknesses Moves down from rank 8
- 2025:LLM08 → 2026:LLM09: Vector and Embedding Weaknesses moves down from rank 8 to 9.