New CamoDocs attack targets RAG models, evades defenses

Researchers have developed a new data poisoning technique called CamoDocs, specifically targeting retrieval-augmented generation (RAG) language models. This method avoids direct query inclusion in poisoned documents, making it harder for existing defenses to detect. CamoDocs synthesizes benign and adversarial content, using dispersion tokens to spread malicious embeddings and coherence filtering to maintain readability. The attack proved effective against proprietary models like GPT-5.4-mini and Claude Haiku 4.5, achieving significant attack success rates while also demonstrating that some defenses, like TrustRAG, can reduce effectiveness but at the cost of utility on benchmarks such as NeoQA. AI

IMPACT This attack highlights a significant vulnerability in RAG systems, potentially impacting the reliability and security of AI applications that rely on external data.

RANK_REASON The cluster contains a research paper detailing a novel attack method against AI systems. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• AI training data poisoning
• CamoDocs
• Claude Haiku 4.5
• GPT 5.4 Mini
• NeoQA
• retrieval-augmented generation
• TrustRAG

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →