Since 2018, a dedicated team within Microsoft has attacked machine learning systems to make them safer. But with the public release of new generative AI tools, the field is already evolving.