AI researcher j⊒nus claims newer Claude models resist prompt manipulation by evaluating user authenticity through costly signals · Digg