Your Model’s Memory Has Been Compromised: Adversarial Hubness in RAG Systems

Cisco researchers have discovered a new attack vector called 'Adversarial Hubness' that compromises the memory systems of RAG (Retrieval-Augmented Generation) AI models. This vulnerability allows attackers to manipulate how AI systems retrieve and process information from their knowledge bases, potentially enabling more sophisticated prompt injections and jailbreaks that bypass existing security measures.