FishCodeTech/muteki

What it solves

Project Muteki is designed to prevent AI agents from falling into "dead-loops"—where a single agent becomes stuck on a specific problem and cannot progress toward a goal. It automates offensive and defensive security tasks, such as Capture The Flag (CTF) competitions, penetration testing, and code auditing, by coordinating a swarm of diverse AI models to work together rather than relying on a single LLM.

How it works

Muteki uses a heterogeneous multi-model architecture where different coding agents (such as Claude, Codex, Cursor, and DeepSeek) collaborate via a shared blackboard. This blackboard acts as a central repository for discovered facts and evidence; when one agent finds a lead, others can use it, and when an agent hits a dead end, others avoid repeating the mistake.

The process follows a specific loop:

1. Coordinator/Reason Planner: Reads the blackboard and plans the next step.
2. Intent: An intent is posted to the blackboard.
3. Workers: Multiple agents claim intents and execute real CLI commands and security tools.
4. Verification: Results are written back to the blackboard, and flags are only accepted if they appear verbatim in the actual execution output.
5. Review: The system periodically reviews recorded facts to correct course and avoid loops.

Who it’s for

Security researchers, penetration testers, and CTF players who want to automate the discovery and exploitation of vulnerabilities using a multi-agent AI swarm.

Highlights

• Multi-Model Swarm: Integrates various engine CLIs including Claude, Codex, Cursor, Pi, OMP, Kimi, Grok, and DeepSeek.
• Proven Performance: Successfully "AK'd" (solved all) challenges in various benchmarks and competitions, including NYU CTF and HackTheBox Insane/Hard categories.
• Autonomous Tooling: Agents can improvise by writing and installing their own dependencies and scripts during a run.
• Flexible Deployment: Supports both local execution and containerized deployment via Docker Compose, with a dedicated Kali Linux-based worker image.