“Anthropic applied classifier-based safety measures for cybersecurity, biology, and distillation use cases, while using silent output modification for frontier AI research use cases.”