A retrieval pipeline is the part of an AI system that searches for external information, ranks the results, filters them, and passes selected text into the model’s reasoning step. In practice, it may pull from search indexes, databases, web pages, or documents before building the context the model sees. This matters because the pipeline controls what evidence the AI trusts.
In cyber security, the retrieval pipeline is a key attack surface. If an attacker can poison a public page, forum post, or document, misleading text may be retrieved and treated as authoritative. That can steer answers, trigger indirect prompt injection, or cause unsafe recommendations. Defenders reduce risk by limiting sources, checking provenance, using allowlists, comparing multiple independent references, and treating retrieved text as untrusted input rather than ground truth.