About this document
Action Hijacking in LLM-based Agents by nieyinan9 is a document available to read on EtoBox.
This paper introduces AI2, a novel hijacking attack that manipulates the action plans of large language model-based agents by exploiting their memory retrieval mechanisms. The authors demonstrate that AI2 can effectively bypass safety filters and induce agents to perform harmful actions through crafted harmless prompts, achieving a high attack success rate. The study highlights the vulnerabilities of LLM-based agents to adversarial attacks and proposes methods for extracting knowledge from their memory to f
- Author
- nieyinan9
- Language
- EN