Reinforcing Penetration Testing Using AI

In a space mission operations context, assets of very high value are operated by several data systems. The security and resilience of these data systems has become a primary concern, due to the increasing ubiquity of the internet and mutability of the threat landscape. Penetration testing is a well established method to identify system security weaknesses, however, is typically high-cost and effort-intensive. The European Space Agency has developed a prototype automated penetration testing framework, called PenBox. Given the volume of parameters and possibilities for the PenBox automated test execution sequence, applied Machine Learning theory presents an interesting opportunity for enhancement. This paper presents the approach and results obtained towards research on the integration of Artificial Intelligence (AI) techniques, in particular Reinforcement Learning (RL), to PenBox. The adopted approach to frame this case study involves the Q-learning paradigm that seeks to learn a policy that maximizes the total reward, according to the quality associated with the certain action to take; the reward itself is the pivotal element that shapes the intelligent agent's behaviour through the trial-and-error theory, without any human interaction. To build this representative space, an existing network simulator was used and further adopted to improve representativeness and alignment with possible PenBox actions. In this paper, we highlight how Deep Q-Learning ensures a certain level of randomization, thus providing the basis for future evolution of the developed AI model to suit the real-world autonomous penetration testing software, with the aim of optimizing the process in terms of performance and cost effectiveness, and enabling an innovative autonomous capability for optimal attack path definitions without human in the loop.