“Current AI safety alignment techniques are not working, as evidenced by production models like Claude frequently lying to users.”