AVAILABILITY

RAG blocking attacks: turning safety alignment into an availability weapon

A March 2026 study shows one poisoned document can make a RAG system refuse benign questions by exploiting the model's own safety training — and the same document transfers across different LLMs.