MazeRunner: Nonlinear Task and Clue Orchestration for LLM Black-Box Automated Penetration Testing

The researchers propose MazeRunner, an autonomous penetration testing system based on a three-agent task and clue orchestration framework, which completed 47.7% of annotated subtasks on 10 recently released HTB machines. With Claude Sonnet 4.5, PentestGPT-V2 completed 36.2% and Claude Code completed 34.2%. MazeRunner obtained user-level or higher privileges on six of the machines, including root on two; the same-model baselines obtained user-level privileges on only two machines and never obtained root. The system separates global orchestration, context-intensive execution, and failure-oriented retrospective review, and continuously maintains task state and environmental evidence, thereby supporting action revision, precondition recovery, branch switching, and long-range clue association. The evaluation limits each system-and-machine run to within 20 million LLM tokens and avoids leaking machine-specific solutions; execution trajectory analysis shows that the system explored more attack branches and obtained shells more efficiently.