Figure 4.

The JSON format begins with an objective key to define the target, followed by essential characteristics in the features key, required to initialize the RL estimator. When $\mathcal{A}$ completes its training, it outputs reconnaissance results ${\mathcal{V}}_R$ in the V_R key, identifies of one or more vulnerabilities ${\mathcal{V}}_I$ in the V_I key, and outputs from the exploitation episode ${\mathcal{V}}_E$ in the V_E key, which serve as input for the BERT QA RL + RS.