RAG Credential Harvesting
This is a security risk where a company's internal AI accidentally learns secrets, like passwords or keys, from private documents. If a hacker gains access to that AI, they can trick it into revealing those hidden secrets.
RAG Credential Harvesting occurs when sensitive authentication data is ingested into a Retrieval-Augmented Generation (RAG) system's vector database. An attacker with access to the LLM interface can craft prompts to retrieve this data, effectively using the AI as a proxy to bypass traditional access controls.
RAG Credential Harvesting is an exploitation technique targeting the data ingestion and retrieval pipeline of a Retrieval-Augmented Generation (RAG) architecture. By exploiting the semantic search capabilities of the vector database, an adversary leverages the LLM's context window to exfiltrate sensitive credentials inadvertently indexed from internal documentation. This attack vector bypasses standard identity and access management (IAM) by utilizing the LLM's authorized retrieval path to surface privileged information that lacks granular, document-level access control enforcement.