| --- |
| license: openrail |
| language: |
| - en |
| size_categories: |
| - 100M<n<1B |
| --- |
| |
| 102,400,000 timesteps, Multi-Agent Reinforcement Learning |
|
|
| Total Environment Steps= 10 parallel environments × 7,000 episodes ×2,048 steps= 102400000 Training Timesteps |
|
|
| -The Red Agent’s goal is to discover vulnerabilities, elevate privileges, compromise assets, and maintain persistence. Its action space can be modeled after phases of the |
| MITRE ATT&CK framework. |
|
|
| -The Blue Agent’s goal is to maintain system availability, reduce the attack surface, detect malicious behavior, deceive the attacker with honeypots and evict the intruder. Its action space |
| consists of defensive configurations and mitigation steps. |
|
|
| Pure Zero-Sum Opposition: Red's gain is Blue's absolute loss and vice versa |
|
|
| Strictly for educational purposes. Do not modify to use against unauthorized Networks |
|
|
| Defense Agent in Action: https://www.youtube.com/watch?v=3g3plOjZkJw |
