>will search for backdoors

Or place new backdoors!

But I still don't see the argument for using closed-weights models. Why should I trust a closed model where I can see none of its weights, activations, or reasoning more than a model where I CAN see its weights, activations, and reasoning?