Friday, Jul 10 | --:--
Back to home

Policy

AI Security China United States

OpenAI, Anthropic, and Google Share Intelligence via Frontier Model Forum to Combat Chinese AI Model Copying

On April 6, 2026, reports emerged that OpenAI, Anthropic, and Google have begun collaborating through the Frontier Model Forum to detect and counter adversarial distillation by Chinese labs, sharing information on attempts to extract capabilities from US frontier models in violation of terms of service.

Tech Insights Reporter 5 min read San Francisco

TLDR\n\nRivals OpenAI, Anthropic, and Google (Alphabet) are sharing threat intelligence through the Frontier Model Forum (FMF)—the industry nonprofit they co-founded with Microsoft in 2023—to detect “adversarial distillation” attempts by Chinese competitors. These efforts involve automated querying of US models to extract outputs and capabilities for training cheaper copycat systems, violating terms of service. OpenAI has accused DeepSeek of such practices; Anthropic has identified DeepSeek, Moonshot AI, and MiniMax as actors, documenting millions of suspicious exchanges.\n\n## Collaboration Details\n\nThe companies are coordinating to identify and block distillation attacks that “free-ride” on US-developed frontier capabilities without incurring equivalent development costs or safety alignments. This mirrors cybersecurity information-sharing practices.\n\n- The effort operates via the FMF, which has previously issued briefs on adversarial distillation (February 23, 2026).\n- OpenAI confirmed participation and referenced prior warnings to Congress about Chinese extraction.\n- Anthropic and Google (via FMF) have been involved in detection efforts.\n- Reports cite 16 million suspicious exchanges documented in some analyses.\n\nThe collaboration comes amid growing concerns over Chinese labs rapidly closing the gap on US models through distillation rather than independent training at frontier scale.\n\n## Context on Distillation\n\nAdversarial distillation uses high-volume, targeted queries to frontier models to harvest reasoning traces, outputs, and capabilities. These are then used to train smaller or competing models that replicate much of the performance at lower cost. US labs view this as both a terms-of-service violation and a national security/IP concern.\n\nChinese models such as those from DeepSeek have shown surprising capabilities relative to claimed training resources, prompting scrutiny.\n\n## Why this story matters\n\nThis marks a rare instance of direct operational cooperation among fierce US competitors on a shared threat. It highlights the intensity of the US-China AI competition, where capability transfer via distillation is seen as a strategic vulnerability. The use of the FMF as an intelligence-sharing channel formalizes what was previously ad-hoc, potentially setting precedents for industry self-defense against model theft and unauthorized replication. It also underscores the economic stakes: distillation allows lower-cost competitors to erode the moats of massive R&D investments by US labs.\n\n## Sources\n- Bloomberg: “OpenAI, Anthropic, Google Unite to Combat Model Copying in China” (April 6, 2026). https://www.bloomberg.com/news/articles/2026-04-06/openai-anthropic-google-unite-to-combat-model-copying-in-china\n- Frontier Model Forum issue brief on Adversarial Distillation (February 23, 2026). https://www.frontiermodelforum.org/issue-briefs/issue-brief-adversarial-distillation/\n- OpenAI congressional memo references on Chinese extraction practices.\n- Additional reporting from The Decoder, Tech in Asia, and others confirming the April 6 disclosures and specific labs named (DeepSeek, Moonshot, Minimax).\n\n## Featured Image Alt Text\n\nLogos of OpenAI, Anthropic, and Google connected by a shield icon representing the Frontier Model Forum, with faint Chinese lab logos in the background symbolizing the collaboration against model distillation announced around April 6, 2026\n\n## Tags\nOpenAI, Anthropic, Google, Frontier Model Forum, Distillation, China, AI Security, Model Theft, Adversarial Extraction

Scroll to continue reading