RAND Europe has published a report warning that attacks designed to fool or disrupt military AI systems could make crises between states harder to control. The study, commissioned by the UK Foreign, Commonwealth and Development Office and published on 29 September, looks at what it calls adversarial AI: attacks that “disrupt, deny or manipulate” military and critical infrastructure systems.
Working through five illustrative scenarios, the authors reach four findings. If allies and rivals believe a military’s AI can be fooled, its deterrence looks less credible. Such attacks may tip the balance towards offence. Because a deliberate attack is hard to tell apart from an ordinary malfunction, and harder still to attribute, there is more room for miscalculation and mistrust. And where conventional and nuclear command systems are entangled, an attack on one could hit the other by accident.
The main recommendation is that governments should treat adversarial AI “as a strategic stability risk, not merely a cyber security or AI assurance issue”. The report also calls for sharing incident and attribution information with allies, building fallback options into AI-enabled military systems, and agreeing red lines and communication channels with competitors.
The report is one of three in a RAND Europe series on strategic stability. Its findings rest on scenarios, not recorded incidents.