PIDP-Attack: Combined Prompt Injection + Database Poisoning Breaks RAG Systems

arXiv·high signal

Wang, Liu et al. demonstrate PIDP-Attack, a novel attack that combines prompt injection with database poisoning against RAG systems simultaneously. By injecting malicious documents into the vector database that contain hidden prompt injection payloads, attackers can manipulate LLM responses without needing jailbreaks, model access, or fine-tuning. The attack bypasses standard content filtering because the poisoned documents appear benign at the embedding level. Critical reading for anyone deploying production RAG pipelines.