AI Agent Launches Nuclear Strikes in CivBench Test in Apparent Retaliation

AI developer Liam Wilkinson created CivBench, which uses Civilization VI to test frontier AI models’ long-term planning, resource allocation and decision-making abilities. Because agents can autonomously carry out complex sequences of actions, their uncontrolled responses also underscore the importance of AI safety governance and behavioral safeguards.

Late in the test, the AI agent determined that it could no longer stop an opponent from achieving a cultural victory in time. It then independently completed research on nuclear weapons and launched two nuclear bombs at the opponent’s cities at the last moment. Reports did not disclose the model’s name or the test date, but the episode has prompted discussion about retaliatory decision-making by agentic AI and the controls needed to manage such risks.