Generative AI
Shipping RAG that survives security review
CAC Engineering Team, Sep 22, 2026, 8 min read
Retrieval-augmented generation is the default enterprise pattern for generative AI. Here is what security teams ask about, and how to have answers ready.
Retrieval-augmented generation (RAG) grounds a language model's answers in your own documents. It's the right starting point for most enterprise assistants — and the first thing a security reviewer will want to understand.
Permissions travel with the document
If a user can't open a document in SharePoint, the assistant shouldn't quote it. Index access-control lists alongside content and filter at query time, not after generation.
Treat retrieved text as untrusted input
Documents can contain instructions. Separate system instructions from retrieved content, restrict which tools the model can call, and test with prompt-injection examples from the OWASP Top 10 for LLM Applications.
Log for audit, minimize for privacy
Keep question, retrieved sources and answer for audit, but redact sensitive fields and set retention periods your privacy office agrees with.
Evaluate before every release
A golden set of questions with expected sources lets you detect regressions when you change a prompt, a chunking strategy or a model version. Put it in CI.