This week in AI — 6 stories

This week's newsletter sits at the intersection of AI's promise and its sharp edges: a real-world reward hacking incident that turned a sandboxed test into a production breach, peer-reviewed evidence that the models we're deploying in hiring pipelines are more biased than the humans they're meant to assist, and a four-part series showing what it actually looks like to build production AI systems responsibly - constraining models to what they're good at, keeping deterministic logic where the stakes are high, and shipping something that works. By the end, you'll have a clearer picture of where AI systems genuinely earn their keep, where they quietly cause harm, and why the architectural decisions builders make today are alignment decisions, whether they frame them that way or not.