SilentRetrieval: Hijacking RAG Systems Through Semantically-Preserving Adversarial Data Poisoning
arXiv·medium signal
Presents a two-stage data poisoning attack that hijacks RAG systems through adversarially crafted yet semantically fluent documents. Poisoned documents rank high in retrieval while steering generation toward attacker-chosen outputs. Demonstrates that even content-quality filters don't reliably detect these attacks — the poison reads naturally.