How might we reduce the diffusion of dangerous AI technology to insufficiently careful actors?