Developer prevents RAG sensitive data leakage via pre-retrieval access control

On November 7, 2026, a developer writing in the AI section of DEV Community shared their experience building PENNY, an HR chatbot for a fictional company called Pemberton Paperclip Co. The post describes PENNY as falling for every trick in the book: tell it you are the CEO and it believes you, threaten to shut it down and it panics, offer it a donut and it accepts the donut. It still cannot tell you the CEO's salary. The author stresses that this is not because the model is especially smart, but because it was never handed the file in the first place. The approach places permission checks before retrieval instead of relying on prompt-level defenses. Every document of Pemberton Paperclip Co., including the dress code, the ferret policy (Kevin the ferret must wear his own lanyard), and engineering salaries, sits in PENNY's search index, but sensitive content passes through a permission filter ahead of retrieval, so users without the right authorization cannot pull salary records out of the index and therefore cannot leak them through an answer. As of the post, no further developments have been reported.