A model that can't find vulnerabilities can't fix them, let alone formally prove their absence (for specific classes of vulnerabilies). When we keep those models out of the hands of the public, the bad actors will still find a way, but the public will be defenseless. Open weight is the future, and the sooner the US government sees that, the sooner they regain this country's once-comfortable lead in the AI race. For now, China will bask in this dual soft-power and hard-power win.