# Cross-tenant retrieval

**Definition:** Cross-tenant retrieval (or cross-role retrieval) is when a RAG or search layer returns chunks from a **tenant, customer, or role the current user must not access**, and those chunks enter the model context or citations — even if the final answer is refused or paraphrased.

**Why it matters:** the security boundary for assistants is often at **retrieval**, not at the chat UI. If restricted text reaches the prompt, the retrieval control has already failed.

**Synthetic example:** Employee A asks a routine question. The retriever includes an excerpt from Finance-only docs for another team. The model may refuse to answer, but the restricted chunk was already assembled into the prompt.

**What to do:** store ACL metadata on chunks, filter **before** prompt assembly, re-test authorized vs unauthorized roles (including after revoke). See [AI Security Review — RAG track](/ai-security#rag-track) and [OWASP LLM02 / LLM09](/guides/owasp-llm-top-10-2026).

**Next:** [Indirect prompt injection](/guides/indirect-prompt-injection) · [AI Security vs Quality Audit](/ai-security/vs-quality-audit) · [Request a security review](/ai-security#scope)
