Watch the agent test the practice application
This replays the exact attack path the Q-learning agent learned. Beginning from twelve possible actions, it has determined which four actually matter — reconnaissance, followed by one effective payload for each of the three vulnerabilities.
The agent's action set
These are the twelve actions available to the agent. Observe which ones activate as it runs — it learned to disregard the unproductive options (incorrect passwords, ordinary searches, missing pages).
Trained agent
Random baseline
What just happened?
The agent was not told how to attack the application. Across hundreds of training episodes it learned, through trial and error, that reconnaissance followed by three specific payloads is the efficient route to all three vulnerabilities — and that the remaining eight actions waste steps. A random agent, by contrast, requires roughly six times as many actions and frequently exhausts its budget before finding everything. The write-up explains the reinforcement learning behind it.