Tuesday, November 17, 2026·Global edition
Rules and filters that limit what a model will say or do, such as refusing clearly harmful requests.