Research · curated 8 Sep 2026
CamoDocs: A Poisoning Attack Against Retrieval-Augmented Language Models Using Camouflaged Documents
First reported arxiv.org
Coverage timeline
Single-source research — first reported, latest, and curated coincide.
Why it matters
CamoDocs shows that RAG poisoning can bypass common embedding- and lexical-based filtering defenses, meaning defenders relying on query-overlap detection or clustering-based erasure may still be vulnerable to targeted answer manipulation.
CamoDocs is a research paper proposing a data-poisoning attack against retrieval-augmented generation (RAG) systems that avoids direct query inclusion by camouflaging adversarial documents among benign content using 'dispersion tokens' and coherence filtering. Across seven RAG defenses, three open-weight LLMs, and three benchmarks, the attack achieves strong attack success rates while evading query-overlap detection, and remains effective against proprietary models (61.80% ASR on GPT-5.4-mini, 55.09% on Claude-Haiku-4.5). Code is published at github.com/jaewonalive/CamoDocs.