TL;DR

OpenAI, Anthropic and Google have started sharing information through the Frontier Model Forum to detect what the industry calls “adversarial distillation” — attempts by Chinese rivals to extract outputs from cutting-edge US models in violation of their terms of service. It is a rare instance of the three biggest US labs cooperating directly against a common competitor.

A Quiet Industry Pact

The three companies founded the Frontier Model Forum with Microsoft in 2023 as a self-regulatory body, but until now its public output has skewed toward safety research and policy positioning. Bloomberg reports that the forum is now also being used as the operational channel for sharing detection signals on Chinese distillation attempts. The trigger, by all accounts, is the rise of Chinese frontier rivals and the suspicion that some of their training data was harvested by automating prompts against US APIs.

This is the second concrete step in a few months toward multilateral self-policing among US labs. It comes alongside individual enforcement actions — Anthropic recently moved to block third-party harness usage on subscriptions, and Google took similar action against unauthorised harvesting of its Gemini CLI authentication earlier this year.

Distillation as a Strategic Risk

The technical concern is that even without access to model weights, a determined competitor can use API outputs as training signal for a smaller model that captures much of the original’s capability. The commercial concern is more direct: it erodes the moat US labs are spending tens of billions on compute to maintain.

Looking Forward

For UK businesses, the practical implication is that frontier model providers are moving toward a more controlled distribution model — tighter terms of service, more aggressive harness restrictions, and greater scrutiny of high-volume API usage. It also strengthens the case for Britain’s alignment with US export controls, since UK customers ultimately depend on the same upstream models the labs are now trying to ringfence. The unresolved question is whether the cooperation extends beyond detection into shared enforcement, or stops at intelligence sharing.