This week in AI — 3 stories

This week's stories share a quiet through-line: the gap between what AI systems can do and what we can safely verify about them is narrowing in some directions and widening in others, and the field is finally being forced to reckon with both. Anthropic's candid bioweapons admission marks a rare moment of institutional honesty about capability ceilings being breached, while researchers on the infrastructure side are pushing those ceilings further still by squeezing dramatically more performance out of already-deployed models. Against that backdrop, a practical paper on deterministic agent oversight offers a grounding reminder that some of our best safety tools don't require more AI at all - just smarter engineering around the AI we already have.