This position paper argues that language models (LMs) are not yet safe for use in strategic conflicts, such as open-ended wargames, without an auditable safety case. The authors identify five failure modes of LM-enabled wargames and propose stress-testing decision-influencing LM agents as a safer approach.