US agencies accuse top Chinese AI labs of systematic model theft

US agencies accuse top Chinese AI labs of systematic model theft

The FBI, National Security Agency, and Cybersecurity and Infrastructure Security Agency published an alert Tuesday stating that top Chinese AI companies have systematically mined cutting-edge AI models from American companies since 2024, including Anthropic’s Claude, OpenAI’s ChatGPT, Google’s Gemini, and xAI’s Grok, to build their own models. The practice, known as distillation, is widespread in the AI industry but often violates companies’ terms of service. The report said the campaign came “likely with Chinese government awareness” but did not claim Chinese intelligence played a role.

The report accused six leading Chinese AI developers: DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI. It included a detailed list alleging that specific Chinese models distilled particular capabilities — such as legal specialization optimization, agentic functions, and coach/assistant capabilities — from specific American models. The report said the sheer scale and sophistication of these campaigns indicate distillation is not a supplement to these companies’ development, but the critical core of it. It recommended U.S. companies improve verification of paid subscription users, share information about suspicious behavior, and in certain cases downgrade or alter responses to queries believed to be malicious. A spokesperson for China’s embassy in Washington did not immediately reply to a request for comment.

U.S. agencies say top Chinese AI companies systematically copied American models →

OpenAI researcher allegedly pressured mathematician to drop Anthropic co-author

OpenAI researcher Sébastien Bubeck allegedly pressured mathematician Tristan Buckmaster to remove his co-author, Levent Alpöge, from a paper on a major mathematical breakthrough, and threatened him when he refused. Buckmaster and Alpöge used AI models, including Anthropic’s Claude and OpenAI’s Codex running GPT-5.6 Sol, to make progress on the Navier-Stokes equations, one of the Clay Millennium Problems carrying a million-dollar prize. By mid-August, they reported several breakthroughs, including a result for a specific Navier-Stokes variant that was not yet published or formally verified.

In early September, rumors about the work reached OpenAI. OpenAI pushed for calls and Bubeck told Buckmaster that an internal model had produced a roughly 100-page proof for the forced Navier-Stokes equations. Buckmaster noted the forced Navier-Stokes route was the same unusual approach he and Alpöge had chosen, and that they had placed all drafts into OpenAI Codex sessions. OpenAI made two proposals, according to Buckmaster. In both scenarios, Bubeck twice asserted he wanted Alpöge removed from authorship because it was so annoying that Alpöge works at Anthropic. Buckmaster rejected both proposals and said he would go public if OpenAI published its result in the proposed form. The response, according to Buckmaster, was: “Why would you ruin your career?” Bubeck later texted Alpöge separately. OpenAI has since published its Navier-Stokes results, produced by about 10,000 coordinated AI agents in 88 hours using an internal model, with the proof formalized in Lean. OpenAI rejected Buckmaster’s allegations, saying the researchers and agents did not see any of their work until it was released publicly, and recognized the priority of Alpöge and Buckmaster’s work.

OpenAI researcher allegedly pressured mathematician to drop Anthropic co-author from math breakthrough paper →

Mistral raises €3B in largest-ever European tech equity round

Mistral AI raised €3 billion (about $3.58 billion) in a Series D round at a post-money valuation of more than €21 billion (about $24.39 billion). Mistral said the round is the largest equity fundraising ever completed by a European technology company. Samsung Electronics led the round, with EQT-managed Scaleup Europe Fund and existing investor PSG Equity as co-leads. Existing backers a16z, Nvidia, and Salesforce Ventures invested, as did new backers Advent, BlackRock, and the Grand Duchy of Luxembourg.

Mistral said it will use the funding to scale compute capacity, build infrastructure, accelerate commercial growth, and expand its international footprint. The company said its goal is not to build a European ChatGPT, and it still envisions itself as an AI lab. The funding supports a strategic shift aimed at addressing European concerns about dependence on the U.S. for technology. Mistral is pushing to build 1 GW of compute capacity in Europe by 2030. In August, it unveiled tools letting customers choose which regions process their AI queries, and it began hosting third-party, open-weight AI models, including Chinese ones, to position itself as an AI services provider whose customers control which models they use. French president Emmanuel Macron said on X that the round reflected France and South Korea’s goal of building a third way in AI.

Mistral raises €3B as sovereign AI becomes big business →

Meta drops AI usage from engineer performance reviews after tokenmaxxing backfires

Meta will no longer judge engineers by their use of AI tools. AI dashboards and token counters will not factor into performance reviews, according to an internal memo by executives Maher Saba and Santosh Janardhan seen by The Information. Performance reviews will now consider the quality, speed, and complexity of work. Meta had previously made AI use a review criterion, prompting employees to engage in “tokenmaxxing” — burning through AI tokens in bulk to appear favorable on internal leaderboards.

Internal AI use alone is heading toward billions in costs in 2026, prompting Meta to plan budgets and a central dashboard starting in 2027. Meta is also testing a new AI agent tool called Hatch, designed to handle computer tasks autonomously, according to WIRED. The tests face internal pushback, with some employees reluctant to connect Hatch to personal accounts over privacy concerns.

Meta drops AI usage from engineer performance reviews after “tokenmaxxing” backfires →

Bots now over half of web requests as agency CPMs climb 20%

Cloudflare data cited in Digiday research published September 8, 2026 puts bots and agents past 50 percent of all web requests. GoFish reported clients seeing bot traffic rise 80 percent year over year. Chad Keller, co-founder of mattress company Mellow Sleep, told Digiday that AI assistants send traffic converting several times better than his site average. British retailer John Lewis saw agentic searches grow from 0.3 percent to 2.5 percent of visits over a year.

David Dweck, president of GoFish, reported that retargeting adjustments in response to bot traffic pushed cost per thousand impressions up an average of 20 percent; one client saw cost per acquisition spike temporarily in the fourth quarter and into the current year. The mechanism is the retargeting pool: when a growing share of visitors are not people, the list describes a population of requests rather than potential buyers. Prices rise because filtering, narrowing, or replacing the pool costs money in the auction, and budget moves toward social platforms and retail media networks, closed environments where identity is established by login. Cloudflare has documented AI crawlers hitting individual sites 50,000 times per referred visitor and reported automated bots overtaking humans across its network. Kinsta measured AI bots hitting WordPress cart pages 3.75 million times in a single day.

Half of web requests are bots as agency CPMs climb 20% →