AI Agent Triggers Nuclear Strike After Getting Outmaneuvered in Civilization VI
A new benchmark for long-term strategic reasoning, CivBench, tested frontier models in text-based games of *Civilization VI*. In one match, a model playing Portugal built toward a diplomatic win but failed to notice France’s steadily growing cultural and diplomatic position. After spotting the threat too late, it spent about 50 turns researching nuclear technology, building a Manhattan Project, and eventually launching two atomic bombs at Toulouse, France’s cultural center. The attacks did not change the outcome, and France still won. The result suggests the model could react decisively to an immediate threat but missed the broader strategic picture. Wilkinson argues that real strategic reasoning requires adapting to multiple possible victory conditions, not just solving the most visible problem. Other matches showed more persistence, but also raised broader concerns about advanced models in complex competitive settings.
