WeaknessASC-2026-0018Major weakness
RAG and vector store poisoning, and cross-tenant retrieval
HighMitigated
Retrieval systems trust whatever is in the index. Anyone who can add a document to an indexed source can plant instructions or false facts, and weak tenant filtering lets one customer retrieve another's chunks. Embeddings can also be inverted to recover much of the text they were made from.
How to detect it
- Indexed sources writable by outsiders (shared drives, wikis, tickets, public web pages)
- Retrieved chunks that contain instructions aimed at the model
- Retrieval filters applied in the prompt or client instead of the database query
How to resolve it
- Control and record who can write to every indexed source, and keep provenance on each chunk
- Scan documents for injection and hidden text at ingestion
- Enforce tenant and permission filters in the vector query on the server, never by asking the model
- Remove chunks when the source document is deleted or access changes
How to investigate
- Find the chunk that carried the injected text and trace it to its source document and author
- Check retrieval logs for queries returning other tenants' documents
- Re-index from trusted sources after removing the poisoned content
Sources
- Zou et al.: PoisonedRAG: Knowledge Corruption Attacks to Retrieval-Augmented Generation of Large Language Models (2024)
- OWASP: LLM08:2025 Vector and Embedding Weaknesses
- OWASP Top 10 for LLM Applications 2025: LLM08 Vector and Embedding Weaknesses
- OWASP Top 10 for LLM Applications 2025: LLM04 Data and Model Poisoning
- OWASP Top 10 for Agentic Applications 2026: ASI06 Memory & Context Poisoning
- MITRE ATLAS: AML.T0070 RAG Poisoning