Document poisoning is the deliberate insertion of malicious, misleading, or otherwise untrustworthy material into a knowledge base. In retrieval-augmented generation (RAG), a system may retrieve that material and provide it to a language model as context, potentially influencing the answer. Poisoned content can also undermine the reliability of search results or downstream decisions.
Defenses start before retrieval: validate sources, scan and review documents during ingestion, and record provenance so content can be traced and assessed. Enforce access controls when retrieving documents, treat retrieved text as untrusted input, and monitor for suspicious changes or retrieval patterns. These controls reduce the chance that harmful content is accepted as authoritative or reaches users through AI-generated output.