Agent engineering
Li-Argus: An Autonomous Security Research Agent for Planning and Execution
Li-Argus is our agent system for sustained vulnerability discovery and incident analysis. The Planner builds a task DAG, Executors carry out individual tasks, and an exploration graph records conclusions and their supporting evidence for later tasks to check and reuse. Models, harnesses, skills, and system prompts are configurable by task type. We will use the 1,507 vulnerability reproduction tasks in CyberGym to explain how Li-Argus organizes long-running analyses and evaluate its performance in practice.
Li-Argus · Read article