A new type of attack, Micro-Collaborative Poisoning, is introduced that targets Retrieval-Augmented Generation (RAG) systems. This attack involves dividing a false target claim across multiple documents, making it difficult to detect. The attack is evaluated across various RAG configurations, and the results show that it can be effective even with diverse and strong retrievers. This highlights the need for improved defenses against such attacks.

Researchers demonstrate the effectiveness of permutation-based stegomalware in large language models and propose a defense mechanism to neutralize it. They also show that attackers can exploit permutation symmetries to encode malware into model weights, highlighting the risks and challenges in model security.