Deep Q-Networks (DQN) suffer from large action spaces (potentially 10^4 possible commands). Most state-of-the-art Autopentest-DRL implementations use due to its stability and sample efficiency. For multi-agent scenarios (e.g., red team vs. blue team), MADDPG (Multi-Agent DDPG) is preferred.
Defenders deploy simple firewalls and IDS alerts. The agent learns to add random delays or route through decoys. autopentest-drl