Research

Phantom: General Backdoor Attacks on Retrieval Augmented Language Generation | ACM Transactions on AI Security and Privacy

Page published

Publication date unknown · First observed: 17 Nov 2026

Coverage timeline

17 Nov 2026acm.orgobserved

Single-source research — one report is available.

Why it matters

Phantom demonstrates how a single attacker-controlled knowledge-base document can enable trigger-dependent manipulation of RAG responses across multiple model families.

Phantom is a two-stage attack method that poisons a RAG knowledge base with one document, optimizing it for retrieval when a naturally occurring trigger appears and for inducing attacker-chosen outputs. The researchers demonstrated attacks on Gemma, Vicuna and Llama, reported transfer to GPT-3.5 Turbo and GPT-4, and demonstrated the method against NVIDIA ChatRTX. Attack objectives include answer refusal, reputation damage, privacy violations and harmful behavior.