Three frontier labs discussed writing their own rulebook

Nothing in your stack changes this week, but the direction of travel is worth one cheap precaution. If pre-release auditing of frontier models becomes a norm, the decisions that matter upstream of you are gating and deprecation: which model versions stay available, and how much notice you get. Keep your provider integration behind a thin interface so swapping a model id is a config change rather than a rewrite, and pin model versions explicitly instead of aliasing to "latest". Record, per feature, which model and version produced which behaviour, and keep the evaluation set you used to sign it off. If an upstream model is withdrawn or altered, that record is the difference between a one-day swap and an open-ended investigation.

Treat this as a press report about meetings, not as a compliance event. Do not rewrite an AI policy, and do not accept a vendor’s reference to "industry standards" in a contract while the term has no defined body behind it. Two questions are worth raising now with counsel and procurement. First, whether any existing agreement already commits you or a supplier to standards that do not yet exist. Second, whether a labs-led body would cover the vendors you actually depend on — a group of three does not bind the rest of the market. Antitrust is a live question here rather than a settled one, so expect the shape of any such body to change before it means anything to your contracts.