“The AI control research agenda proposes using red teaming to estimate how difficult it is for a policy to subvert a control technique.”