Behavioural Stealth in Autonomous Red Teaming

Autonomous systems can complete individual security tasks while producing behaviour that is poorly suited to a red teaming exercise. This becomes particularly apparent when several agents are operating concurrently, as local decisions can accumulate to form a pattern of activity that is inconsistent with the environment or the intended adversary. Therefore, behavioural stealth raises a question beyond task completion: how should the collective behaviour of an autonomous red team be understood and assessed from the defender’s perspective?

In this talk, I will examine this question through our work at the Cyber-Defence Campus and ETH Zürich on agentic systems for penetration testing and red teaming automation. I will present the systems we have developed and discuss how memory, coordination and interpretation of results affect their behaviour in sustained exercises. I will also discuss our ongoing research into behavioural constraints in multi-agent cyber operations, focusing on the relationship between local agent decisions and collective activity as observed by defenders. This raises questions about what constitutes plausible adversary behaviour and how that behaviour should be assessed across a sustained exercise.