Steering LLM Viewpoints through Fabricated Evidence Injection

arXiv·medium signal

This paper probes a cognitive vulnerability: LLMs can be pushed toward misleading conclusions when fabricated 'evidence' is injected into context, shifting their stated viewpoint. As chatbots increasingly shape user decisions, the attack shows how planted false context — not jailbreaks — can steer outputs. Important threat model for RAG and agent pipelines that ingest untrusted retrieved content. Published 2026-06-04 (cs.CL/cs.CR).