Frontier LLMs Show Extreme Policies in Idealised AI Development Races

A recent preprint available on arXiv (2608.01193) examines the strategic safety practices of large language models (LLMs) within the context of an AI development competition, structured as a repeated game involving two to five participants. This research highlights a multi-agent safety dilemma where firms face the choice of progressing cautiously and safely or accelerating their development at the risk of forfeiting their ultimate reward. The authors implement an audit gate prior to behavioral analysis, ensuring verification of the game engine, testing rule recall, state tracking, payoff calculations, and stability across various equivalent task descriptions. They then analyze LLM action sequences against an evolutionary game-theory standard and existing human data, investigating variations across models, risk scenarios, personas, and competition sizes. The audit indicates that robust rule recall may exist alongside poor state tracking and payoff calculations, with verified arithmetic potentially affecting results. The results imply that cutting-edge LLMs adopt extreme strategies in these theoretical competitions, contrasting with human behavior.