How microsoft/cyberbattlesim works
CyberBattleSim is a research simulator for training and comparing automated agents on abstract enterprise-network attacks and defense. The main workflow is to select or generate a network scenario, run an attacker through the Gym environment, resolve actions against the simulated network, optionally let a defender detect and mitigate attacks, and use observations and rewards to train or evaluate agents. The repository also provides graph/observation wrappers and benchmark plotting. Connections among unsampled scenario and algorithm modules are limited to documented references; those implementations were not inspected.
Main entry points: cyberbattle_env.py Gym Environment · learner.py Learning Loop · model.py Network Model · agent_wrapper.py Agent Wrapper
Components 25
Simulation Core
- Gym Environment
cyberbattle/_env/cyberbattle_env.py - Network Model
cyberbattle/simulation/model.py - Attacker Actions
cyberbattle/simulation/actions.py - Command Control
cyberbattle/simulation/commandcontrol.py - Stochastic Defender
cyberbattle/_env/defender.py
Network Scenarios
- Environment Generation
cyberbattle/simulation/environment_generation.py - Random Networks
cyberbattle/simulation/generate_network.py - Active Directory
cyberbattle/_env/active_directory.py - Chain Scenario
cyberbattle/_env/cyberbattle_chain.py - Random Scenario
cyberbattle/_env/cyberbattle_random.py - Tiny Scenario
cyberbattle/_env/cyberbattle_tiny.py - Toy CTF
cyberbattle/_env/cyberbattle_toyctf.py
Agents and Training
- Training Runner
cyberbattle/agents/baseline/run.py - Learning Loop
cyberbattle/agents/baseline/learner.py - Deep Q Agent
cyberbattle/agents/baseline/agent_dql.py - Tabular Q Agent
cyberbattle/agents/baseline/agent_tabularqlearning.py - Credential Agent
cyberbattle/agents/baseline/agent_randomcredlookup.py - Random Agent
cyberbattle/agents/random_agent.py - Training Plots
cyberbattle/agents/baseline/plotting.py
Observation Tools
- Agent Wrapper
cyberbattle/agents/baseline/agent_wrapper.py - Graph Wrapper
cyberbattle/_env/graph_wrapper.py - Graph Spaces
cyberbattle/_env/graph_spaces.py - Flatten Wrapper
cyberbattle/_env/flatten_wrapper.py - Option Wrapper
cyberbattle/_env/option_wrapper.py
Other components
- Researcher
Connections 32
- Researcher → Training Runner: starts training
README.md - Training Runner → Gym Environment: configures episodes
cyberbattle/agents/baseline/run.py - Training Runner → Learning Loop: runs learning
cyberbattle/agents/baseline/run.py - Learning Loop → Gym Environment: steps environment
cyberbattle/agents/baseline/learner.py - Learning Loop → Agent Wrapper: uses features
cyberbattle/agents/baseline/learner.py - Learning Loop → Training Plots: records metrics
cyberbattle/agents/baseline/learner.py - Deep Q Agent → Learning Loop: implements policy
cyberbattle/agents/baseline/agent_dql.py - Deep Q Agent → Agent Wrapper: uses features
cyberbattle/agents/baseline/agent_dql.py - Tabular Q Agent → Learning Loop: implements policy
cyberbattle/agents/baseline/agent_tabularqlearning.py - Tabular Q Agent → Agent Wrapper: uses features
cyberbattle/agents/baseline/agent_tabularqlearning.py - Credential Agent → Learning Loop: implements policy
cyberbattle/agents/baseline/agent_randomcredlookup.py - Credential Agent → Agent Wrapper: reads observations
cyberbattle/agents/baseline/agent_randomcredlookup.py - Random Agent → Gym Environment: samples actions
cyberbattle/agents/random_agent.py - Gym Environment → Network Model: uses state model
cyberbattle/_env/cyberbattle_env.py - Gym Environment → Attacker Actions: dispatches actions
cyberbattle/_env/cyberbattle_env.py - Gym Environment → Command Control: uses control logic
cyberbattle/_env/cyberbattle_env.py - Gym Environment → Stochastic Defender: steps defender
cyberbattle/_env/cyberbattle_env.py - Attacker Actions → Network Model: updates network
cyberbattle/simulation/actions.py - Command Control → Network Model: uses state model
cyberbattle/simulation/commandcontrol.py - Stochastic Defender → Network Model: scans nodes
cyberbattle/_env/defender.py - Stochastic Defender → Attacker Actions: uses mitigation actions
cyberbattle/_env/defender.py - Environment Generation → Network Model: builds environments
cyberbattle/simulation/environment_generation.py - Random Networks → Network Model: builds random networks
cyberbattle/simulation/generate_network.py - Random Scenario → Random Networks: creates topology
cyberbattle/_env/cyberbattle_random.py - Random Scenario → Gym Environment: provides environment
cyberbattle/_env/cyberbattle_random.py - Active Directory → Gym Environment: provides environment
cyberbattle/_env/active_directory.py - Chain Scenario → Gym Environment: provides environment
cyberbattle/_env/cyberbattle_chain.py - Tiny Scenario → Gym Environment: provides environment
- Toy CTF → Gym Environment: provides environment
- Graph Wrapper → Graph Spaces: uses spaces
cyberbattle/_env/graph_wrapper.py - Flatten Wrapper → Gym Environment: wraps environment
cyberbattle/_env/flatten_wrapper.py - Option Wrapper → Gym Environment: wraps environment
cyberbattle/_env/option_wrapper.py