Gather RAG-Indexed Targets
Gather RAG-Indexed Targets is the process where a hacker figures out exactly where an AI gets its information. Think of it like a spy watching a library to see which specific books a student uses to study; once the spy knows the books, they can sneak in and swap the pages with false information to trick the student.
This refers to the reconnaissance phase where an adversary maps the external data repositories integrated into a RAG pipeline. By analyzing system documentation or performing targeted queries to elicit source citations, an attacker identifies the specific databases or document stores that inform the model's output, creating a roadmap for data poisoning or supply chain manipulation.
Gather RAG-Indexed Targets describes the adversarial reconnaissance of the retrieval component within a Retrieval-Augmented Generation architecture to identify the provenance of external context. By leveraging side-channel information, system-disclosed metadata, or inference-based probing to map the vector database or document corpus, an adversary establishes a target surface for subsequent data poisoning, integrity attacks, or unauthorized data exfiltration.