Jiawen Shi (Huazhong University of Science and Technology), Zenghui Yuan (Huazhong University of Science and Technology), Guiyao Tie (Huazhong University of Science and Technology), Pan Zhou (Huazhong University of Science and Technology), Neil Gong (Duke University), Lichao Sun (Lehigh University)

Tool selection is a key component of LLM agents. A popular approach follows a two-step process - emph{retrieval} and emph{selection} - to pick the most appropriate tool from a tool library for a given task. In this work, we introduce textit{ToolHijacker}, a novel prompt injection attack targeting tool selection in no-box scenarios. ToolHijacker injects a malicious tool document into the tool library to manipulate the LLM agent's tool selection process, compelling it to consistently choose the attacker's malicious tool for an attacker-chosen target task. Specifically, we formulate the crafting of such tool documents as an optimization problem and propose a two-phase optimization strategy to solve it. Our extensive experimental evaluation shows that ToolHijacker is highly effective, significantly outperforming existing manual-based and automated prompt injection attacks when applied to tool selection. Moreover, we explore various defenses, including prevention-based defenses (StruQ and SecAlign) and detection-based defenses (known-answer detection, DataSentinel, perplexity detection, and perplexity windowed detection). Our experimental results indicate that these defenses are insufficient, highlighting the urgent need for developing new defense strategies.

View More Papers

Risk Assessment for ML-Based Applications in Satellite Systems

Simon Shigol, Roy Peled, Avishag Shapira, Yuval Elovici, Asaf Shabtai (Ben Gurion University of the Negev)

Read More

Automating Function-Level TARA for Automotive Full-Lifecycle Security

Yuqiao Yang (University of Electronic Science and Technology of China), Yongzhao Zhang (University of Electronic Science and Technology of China), Wenhao Liu (GoGoByte Technology), Jun Li (GoGoByte Technology), Pengtao Shi (GoGoByte Technology), DingYu Zhong (University of Electronic Science and Technology of China), Jie Yang (University of Electronic Science and Technology of China), Ting Chen (University…

Read More

Analysing Privacy Risks in Children’s Educational Apps in Australia

Sicheng Jin (University of New South Wales), Rahat Masood (University of New South Wales), Jung-Sook Lee (University of New South Wales), Hye-Young (Helen) Paik (University of New South Wales)

Read More