Autonomous AI Swarm Legion Wins Tencent Intelligent Attack-Defense Challenge

Overview

An autonomous multi-agent orchestration harness developed by open-source contributors has secured first place in the online qualifier of the 3rd Tencent Intelligent Attack-Defense Challenge. Operating fully unattended for eight hours against 634 competing teams, the system utilized a structured AI legion architecture combining an AI commander, worker swarms, and a layered fact-intent graph framework.

According to project documentation, the framework achieved identical scoring results to the runner-up while utilizing significantly fewer tokens, highlighting efficiency in multi-agent task distribution and resource allocation.

Architecture and Core Design

Unlike single-agent models, the project—designated as *sapper-v2*—functions as a coordinated force using the "Nüwa Foundation" base layer. This architecture incorporates three foundational capabilities:

• Create: Spawning processes and agents.
• Language: Enabling inter-agent communication and signaling.
• History: Maintaining a layered fact-intent-recall graph for persistent external memory.

The runtime environment separates responsibilities into distinct operational windows:

• The Commander: Reads the operational board, makes strategic decisions, and dispatches tasks without directly executing attacks.
• The AI Legion: Generals and soldiers execute parallel operations within managed panes.
• The Watchdog: A background script focused entirely on system resilience, managing crashes, key rotation, and session continuity.

Practical Deployment and Verification

Project maintainers emphasize that the framework is designed for competitive scenarios and general long-horizon cooperative work. Deploying the system outside standard evaluation platforms requires proper environment configuration, including setting benchmark base URLs, model gateways, and API keys as detailed in the repository documentation. Verification of official announcements and updates can be conducted directly through the project's official GitHub repository.

Safety and Responsibility

Recognizing the potential risks associated with automated offensive tooling, the authors propose a taggant mechanism. This proposed system aims to embed harmless operational markings during authorized testing while aiding attribution in unauthorized contexts, inviting community collaboration on cryptographic safety designs.