Fabien Roger comments on AI Control: Improving Safety Despite Intentional Subversion