SAPPER-v2 Beats 634 Teams in Hacking Contest Using AI Agents Alone
SAPPER-v2 is an open-source multi-agent pentest framework with one AI commander dispatching a worker swarm ( repo ) Won 3rd Tencent Intelligent Attack-Defense online qualifier, 80.18/100, beating 633 other teams running 8 hours unattended Used 1B tokens to tie the runner-up's 85 flags, who burned
Takeaways
- SAPPER-v2 is an open-source multi-agent pentest framework with one AI commander dispatching a worker swarm ( repo )
- Won 3rd Tencent Intelligent Attack-Defense online qualifier, 80.18/100, beating 633 other teams running 8 hours unattended
- Used 1B tokens to tie the runner-up's 85 flags, who burned 4.5B tokens, via model tiering
- Commander runs GLM-5.3, soldiers run DeepSeek V4.1 Flash; architecture is model and CLI agnostic
- Shared fact-intent graph ( nvwabase ) lets agents die while memory survives for successors
- MIT licensed, ships Docker image with full Kali toolchain; authorized testing only
SAPPER-v2 is an MIT-licensed orchestration framework that divides offensive-security work among a hierarchy of AI agents. A commander reads the competition board and delegates tasks, generals coordinate groups, soldiers execute commands and attempt exploits, and a watchdog restores failed sessions. Shared memory and role-specific models keep the system operating across an eight-hour campaign.
Eight hours, 85 flags
The project’s published results place SAPPER-v2 first among 634 teams in the online qualifier for the third Tencent Cloud Intelligent Attack-Defense Challenge. It scored 80.18 out of 100 while operating without human intervention for eight hours. The system captured 85 flags, the proof strings awarded for solving challenges, and reportedly reached that total 25 seconds before the runner-up.
| Measure | Result |
|---|---|
| Finish | First among 634 teams |
| Score | 80.18 out of 100 |
| Flags | 85, matching the runner-up |
| Time | Reached 85 flags 25 seconds earlier |
| Model usage | 1 billion tokens, compared with 4.5 billion for the runner-up |
The qualifier measured performance in a structured capture-the-flag environment with known scoring rules and a fixed deadline. The result demonstrates sustained coordination under those conditions. Production networks and less constrained penetration tests require separate evaluation.
A hierarchy built to recover
SAPPER-v2 runs its agents in tmux, a terminal multiplexer that maintains several persistent shell sessions. Its runtime separates planning, execution, and recovery into three operational layers:
- Commander: inspects the board, prioritizes targets, assigns work, and delegates attack execution.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
