Resource Hijacking: New Attack Vector on LLM Agents

A new and serious security issue has been identified in large language model agents for the first time. Known as "agent resource hijacking," this flaw takes advantage of important assets that these agents can reach, such as computing power, credentials, budgets, identities, and communication methods. Unlike previous studies that focused mainly on how to attack instructions or data, this research highlights a vulnerability where attackers can manipulate agents to access or control these resources without needing direct access themselves. To assist in further investigation, the researchers have developed ResourceHijackBench, which creates automated scenarios for resource hijacking, organizing valuable resources into six categories and generating 300 attack scenarios with 900 test cases. The findings are available on arXiv under ID 2608.15108, underscoring the urgent need to enhance security for LLM agents as they become more integrated into critical systems.