Abstract
This CETaS Briefing Paper was commissioned by the UK National Cyber Security Centre (NCSC) and explores different views on the authorisation limits for autonomous agents in network defence, including the contexts within which certain threat containment, eradication and recovery actions should be permissible. This study found that introducing autonomous agents in network defence could introduce potential error risks for almost all tasks on the MITRE D3FEND framework and that the desirable level of autonomy for most contexts in cyber defence converge around systems where autonomous agents can carry out operator initiated pre-set tasks independently within specific conditions under supervision. The study team also found that organisation-specific, context-informed and risk-based analysis is required to determine tailored permissions specifications for autonomous agents that reduce the risk to as low as reasonably practical. Finally, the research investigated the thresholds for ‘net benefit’ and ‘sufficient confidence’ to deploy an AI-based cyber defence system, even if all error risks are not eliminated. Achieving such confidence would require autonomous agents to exhibit a degree of context-specific judgement, and increased reliability compared to current human-led systems – for instance through measurable improvement on cybersecurity outcomes or better understanding of edge cases after the introduction of autonomous agents.