Steerability of Autonomous Cyber-Defense Agents by Meta-Attackers

AI agents are increasingly automating several traditional manual tasks. One area where AI agents show promise is computer incident response, as this is a significantly slow and sometimes tedious process managed by operators who are overwhelmed with alarms. However, before deploying these AI systems, we need to make sure that an attacker cannot exploit them. This paper formally analyzes an attacker who has partially compromised an autonomous cyber defense agent. Our goal is to understand how such an attacker can steer the defenses and manipulate the system to achieve its objectives. Our results can help defenders identify the most critical components of their defenses and harden resources that (if compromised) may give an attacker a large advantage.