“Publishing AI safety cases allows thousands of academics, ML researchers, and rival company employees to find flawed assumptions that a small, overworked internal team might miss.”