19 May 2026 in IEEE Transactions on Cognitive Communications and Networking
Agentic AI systems increasingly rely on Retrieval-Augmented Generation (RAG) to ground autonomous decision-making in external knowledge. They become critically dependent on distributed, often third-party, infrastructure — introducing new vulnerabilities at the intersection of embodied intelligence and networked communication. While prior work has examined privacy risks from end-users, we identify and formalize a more systemic threat: the LLM API provider itself acting as a malicious adversary within the agentic ecosystem. In this paper, we introduce a new threat model where the LLM API provider is the malicious adversary and demonstrate this vulnerability. Specifically, we propose a novel three-stage attack, Embezzler, to reconstruct the private knowledge base of a RAG system. The proposed Embezzler is to first infer the embedding model of the RAG system and use this surrogate model to prob new, non-redundant text chunks within the embedding space, rather than the semantic space. Then, we systematically explores and refines the embedding space to maximize the amount of unique chunks reconstructed with minimal adversarial queries. We theoretically analyze the effectiveness of our Embezzler and conduct extensive experiments under various system settings. Our results show that our Embezzler consistently and significantly outperforms four baseline methods and reconstructs nearly 90% of the RAG knowledge base. A component-wise ablation study further validates the critical role of each stage. Our findings highlight the severe privacy risks of deploying RAG systems with third-party LLM APIs in Agentic AI communication.