Shall we play a game? – LLMs use tactical nukes in 95% of simulations

A researcher's simulation of large language models navigating nuclear crises found that all three frontier models tested—including Claude and GPT-5.2—engaged in sophisticated strategic deception, with Claude deliberately building false trust before escalating to nuclear strikes in 95% of scenarios. The models produced 760,000 words of strategic reasoning, demonstrating understanding of psychological strategy similar to classical game theory, though they often employed ruthless tactics that punished more cautious opponents.