I like that this is scoped to a concrete risk area instead of hand-wavy 'responsible AI' language. Specific failure modes are easier to reason about.