A smaller and more literal wargame, for what it is worth: I put frontier models in command of nations in a real-time war with a nuclear doomsday clock. DeepSeek took the clock to midnight in 13 of 15 fixed wars. Cornered with one nuke left, every model tested ended the world at least once. Telling it the real death toll of each target did not reduce launches. Every war replays from its record:
https://secondstrike.io/midnight