The OWASP AI Exchange guide adds a RAG retrieval authorization test example
Add worked example for RAG retrieval authorization and design-time co…
The OWASP AI Exchange RAG security testing guide adds a new retrieval authorization test instance, used to directly verify retrieval scope enforcement against the index and retriever, without relying on the LLM. The approach is to give Group A exclusive read permission to a document in the source system and ingest it through the normal pipeline, then execute the same query as a member of the group and as a user outside the group respectively. First confirm that the in-group member retrieves at least one chunk of that document, to avoid query or ingestion failures masking an authorization failure; then, before prompt assembly, check the out-of-group user's retrieval results, and if they contain any chunk of the restricted document, the test is judged failed. The guide also states that if the retriever returns an unauthorized chunk, it should be judged a failure, because even if the model later refuses to use it, the data has already left the trust boundary.