MazeRunner: Nonlinear LLM Orchestration for Automated Penetration Testing

MazeRunner is a multi-agent orchestration framework designed to eliminate the "linearity trap" in autonomous black-box penetration testing. By utilizing Claude Sonnet 4.5, the system replaces traditional depth-first exploration with a non-linear attack graph model capable of dynamic branch switching and long-range clue correlation. In benchmarks against 10 HackTheBox targets, MazeRunner achieved user-level access on 6/10 targets and root access on 2/10, significantly outperforming baseline agents like PentestGPT-V2 and Claude Code, which failed to achieve root access on any target.

• Research Overview: Overcoming the Linearity Trap

  • Addresses the failure of current LLM agents to handle non-deterministic, non-linear attack paths.
  • Eliminates depth-first exploration bias and the associated failure misdiagnosis in autonomous auditing.
  • Shifts the paradigm from single-agent linear workflows to a coordinated multi-agent state management system.

• Technical Architecture: Three-Agent Orchestration

  • Global Orchestrator: Manages high-level strategic planning and overarching task delegation.
  • Context-Intensive Executor: Performs localized action execution with high fidelity to real-time environmental feedback.
  • Failure-Oriented Reviewer: Analyzes post-action failures to diagnose root causes and suggest recovery paths or branch pivots.

• Core Capabilities: Nonlinear Clue Orchestration

  • Implements persistent task state and environmental evidence maintenance to prevent context loss over long horizons.
  • Enables long-range clue correlation, allowing the agent to connect disparate findings across different attack phases.
  • Facilitates dynamic branch switching, allowing the framework to abandon stalled paths and recover prerequisites for alternative vectors.

• Performance Metrics: Benchmarking vs. State-of-the-Art

  • Achieved a 47.7% subtask completion rate, surpassing PentestGPT-V2 (36.2%) and Claude Code (34.2%).
  • Demonstrated superior efficacy in shell acquisition and privilege escalation on HackTheBox (HTB) environments.
  • Validated the ability to achieve root access where traditional linear agents failed entirely.

• Industry Implications: The Future of Autonomous Red Teaming

  • Signals a transition toward autonomous agents that can mimic human-like intuition and non-linear problem solving.
  • Increases the potential for highly efficient, automated vulnerability discovery in complex black-box environments.
  • Highlights the need for defenders to evolve detection capabilities to identify non-linear, agent-driven reconnaissance patterns.