[bryan0]

The point is that even though meta “conducted extensive red teaming exercises with external and internal experts to stress test the models” a simple attack like this is still possible.