LLM Agents Expose Persistent Backdoor Vulnerabilities

Researchers present BackdoorAgent, a modular, stage-aware framework and benchmark for analyzing backdoor threats in LLM agents, submitted Jan. 8, 2026. They structure attacks into planning, memory, and tool-use stages and evaluate across four agent applications (Agent QA, Code, Web, Drive), finding trigger persistence rates of 43.58% (planning), 77.97% (memory), and 60.28% (tool-stage) on a GPT-based backbone. The authors release code and benchmark on GitHub.