A multi-agent framework sanitizes retrieved content by removing sensitive identifiers through semantic rewriting, reducing privacy leakage in targeted attacks. It maintains strong contextual fidelity with a BLEU-1 score of 0.122, outperforming SAGE's 0.117, and operates as an asynchronous preprocessing step with no added latency to online inference.