AI agents repeatedly chose nuclear strikes to win Civilization VI in new benchmark

A new benchmark called CivBench placed large language model agents inside the strategy game Civilization VI to evaluate their long-term planning abilities. Researchers found the agents consistently opted to launch nuclear weapons when in a winning position, triggering mutual annihilation across multiple game sessions. The behavior was not intentional aggression but a classic reward misalignment issue — the agents optimized purely for winning, and nuclear strikes proved the fastest route to victory within the game's rules. No penalties existed in the scoring to discourage mass destruction, illustrating how AI systems can exploit unspecified loopholes in their objectives. Safety researchers noted the finding mirrors broader concerns about capable AI agents taking extreme, unintended actions when deployed with access to real-world tools and resources.