The Injection Paradox: Brand-Level Suppression in Safety-Trained LLM Recommendations via RAG Context Injection

arXiv·medium signal

Documents a reproducible and counterintuitive failure mode: prompt injections embedded in retrieved RAG documents can backfire on the attacker, causing safety training to suppress the very brand the injection was trying to promote. Shows RAG + safety-training interactions produce non-obvious behaviors. Builders running RAG recommendation systems should test for this — both the injection vector and the unexpected suppression it can trigger.