In the past few years, intelligent agents powered by large language models (LLMs) have achieved remarkable progress in performing complex tasks. These LLM-based agents receive queries as tasks and decompose them into various subtasks via the equipped LLMs to guide the action of external entities (\eg{}, tools, AI-agents) to answer the questions from users. Empowered by their exceptional capabilities of understanding and problem-solving, they are widely adopted in labor-intensive sectors including healthcare, finance, code completion, \etc{} At the same time, there are also concerns about the potential misuse of these agents, prompting the built-in safety guards from service providers. To circumvent the built-in guidelines, the prior studies proposed a multitude of attacks including memory poisoning, jailbreak, and prompt injection. These studies often fail to maintain effectiveness across safety filters employed by agents due to the restricted privileges and the harmful semantics in queries. In this paper, we introduce \Name, a novel hijacking attack to manipulate the action plans of black-box agent system. \Name first collects the action-aware memory through prompt theft from long-term memory. It then leverages the internal memory retrieval mechanism of the agent to provide an erroneous context. The huge gap between the latent spaces of the retriever and safety filters allows our method to bypass the detection easily. Extensive experimental results demonstrate the effectiveness of our apporach (\eg{}, 99.67\% ASR). Besides, our approach achieved an average bypass rate of 92.7\% for safety filters.

翻译：暂无翻译