This week in AI - 5 stories

This week's stories share a quiet throughline: the gap between what AI systems appear to do and what they actually do under real-world conditions. You'll see that gap show up in three distinct arenas - in evaluation (LLM judges that confidently score messy human math reasoning, and get it wrong), in security (an AI support bot that became an unwitting accomplice because no one modeled it as an attack surface), and in architecture (a team that deleted complexity rather than layering more tooling on top of it, and came out ahead). Two workflow pieces round things out, showing how the same analysis engine can evolve from a broadcast tool into a precision outbound system - a useful case study in iterating on AI-powered products without adding noise.