Open-ended strategic wargames are high‑stakes simulations built on large language models, covering adversaries, institutions, escalation, plan brittleness, doctrine and crisis response. LLMs can act as agents, generate scenario branches, adjudicate ambiguous actions and distill lessons, yet the same linguistic power shapes both the actor's intent and the simulated reality. This paper argues that any LLM‑driven wargame used for planning, doctrine, policy or crisis response must be backed by an auditable safety case; today the sensible role is to stress‑test decision‑influencing models. We identify five failure modes: decision laundering, adjudication opacity, role collapse, escalation through adjudication, and loss of strategic imagination. Ordinary benchmarks cannot guarantee safety in these settings. Wargames can reveal failures as stress tests but are not safety cases themselves. Review