This is the repository containing the code and data for the paper Watch Out for Your Agents! Investigating Backdoor Threats to LLM-Based Agents [pdf]
We have released the poisoned training data used in Web Shopping (put in here) and Tool Learning (download from here) experiments. Our training and evaluation code is based on AgentTuning and ToolBench, and users can adopt the code from above two platforms to implement agent backdoor attacks. We are re-organizing the local code, and will release the code when it is ready.