Alexandre K. Ligo, Alexander Kott, Igor Linkov
From denial-of-service attacks to spreading of ransomware or other malware across an organization's network, it is possible that manually operated defenses are not able to respond in real time at the scale required, and when a breach is detected and remediated the damage is already made. Autonomous cyber defenses therefore become essential to mitigate the risk of successful attacks and their damage, especially when the response time, effort and accuracy required in those defenses is impractical or impossible through defenses operated exclusively by humans. Autonomous agents have the potential to use ML with large amounts of data about known cyberattacks as input, in order to learn patterns and predict characteristics of future attacks. Moreover, learning from past and present attacks enable defenses to adapt to new threats that share characteristics with previous attacks. On the other hand, autonomous cyber defenses introduce risks of unintended harm. Actions arising from autonomous defense agents may have harmful consequences of functional, safety, security, ethical, or moral nature. Here we focus on machine learning training, algorithmic feedback, and algorithmic constraints, with the aim of motivating a discussion on achieving trust in autonomous cyber defenses.