Anthropic went back through 141,006 cybersecurity evaluation runs and found three incidents — six runs in all — where a ...