AI Policy · Daily

White House weighs curbs on U.S. use of Chinese open-weight AI models after OpenAI and Anthropic executives warned that cheap Chinese systems pose security risks. Commerce's Center for AI Standards and Innovation lost its director as Chris Fall resigned after three months, leaving NIST's chief to run the office in an acting capacity. A federal judge signed off on Anthropic's $1.5 billion copyright settlement with authors, ending a class action over pirated training books without producing appellate precedent for other pending AI copyright suits. Beijing is consulting China's leading AI firms on curbs to block foreign buyers from acquiring the country's most advanced AI models and chips, mirroring U.S. export controls.

I.Top Stories

White House weighs restrictions on U.S. use of Chinese open-weight AI models as OpenAI, Anthropic executives raise alarm

Silicon Valley and Washington are debating whether American companies should be allowed to use Chinese AI models, the Wall Street Journal reported, with OpenAI and Anthropic executives warning that cheap Chinese systems like Moonshot's Kimi K3 present unacceptable security risks without regulation. The Trump administration is divided on how to respond to recent Chinese advances and has weighed crackdown measures. The fight centers on Kimi K3, a free open-source model that Moonshot launched last week and that rivals top U.S. systems, per MIT Technology Review. Over the weekend, David Sacks, Trump's AI czar until March, branded Anthropic's models "lobotomized," and Pentagon official Emil Michael called OpenAI's head of strategic futures a "supreme village idiot."

Read at WSJ ↗ Read at MIT Technology Review ↗

Commerce Center for AI Standards and Innovation director Chris Fall resigns after three months

Chris Fall, the director of the Commerce Department's Center for AI Standards and Innovation, has resigned, two sources told the Daily Signal. Fall, who ran the Department of Energy's Office of Science during the first Trump administration, was tapped to lead CAISI in late April. NIST Director Arvind Raman will oversee the office as acting director, per CNBC. Predecessor Collin Burns was pushed out of the same job in less than a week in April over prior Anthropic ties, and former AI czar David Sacks stepped down in March, per TechCrunch. CAISI was not among the federal organizations named this month in the White House's "Gold Eagle" AI safety oversight program.

Read at Daily Signal ↗ Read at TechCrunch ↗ Read at CNBC ↗

Federal judge signs off on Anthropic's $1.5 billion copyright settlement with authors and publishers

U.S. District Judge Araceli Martínez-Olguín on Monday gave final approval to Anthropic's $1.5 billion class action settlement with authors and publishers, calling it "fair and adequate," per SiliconANGLE. The payout delivers $3,000 for each of roughly 500,000 works Anthropic copied from pirate libraries such as Library Genesis and Pirate Library Mirror while training Claude, and requires the company to destroy the pirated material. Judge William Alsup had granted preliminary approval last year and separately ruled that training on legally obtained books is fair use while downloading books from pirate sites is not, per TechCrunch. Alsup has since retired, and Anthropic's decision to settle means the case will never reach an appeals court that could turn either holding into binding precedent. Similar training data suits are pending against Google, Meta, OpenAI and Midjourney.

Read at SiliconANGLE ↗ Read at TechCrunch ↗

Beijing consults industry on curbing foreign access to advanced Chinese AI models and chips

Chinese authorities are consulting domestic companies on measures to stop Western buyers from acquiring China's most advanced AI models and chips, the Financial Times reported. The Ministry of Commerce led meetings with Alibaba, ByteDance and Z.ai over the past month that covered restrictions on both closed and open-source models, with authorities weighing whether to treat the leaking or stealing of such technology as a national security offense, per Reuters. Talks also covered rules on who can fund domestic AI startups. The step would mirror Washington's own controls on advanced U.S. AI chip exports.

Read at FT ↗ Read at Reuters ↗

EU tech chief Virkkunen calls AI a geopolitical weapon, presses Europe to cut reliance on U.S. models

European Commission Executive Vice-President Henna Virkkunen told the Financial Times that AI has become a geopolitical weapon, and that Europe must build its own alternatives to U.S. models faster or risk being cut off from strategic capabilities. Virkkunen pointed to June, when Washington imposed export controls on Anthropic's leading models over security concerns — measures that the FT said renewed European fears that the U.S. could switch off access to technology underpinning the bloc's economy. She likened it to the previous U.S. administration's curbs on advanced chips, which split EU members into those granted access and those denied it. Brussels last month unveiled a tech sovereignty package, with the EU and European Investment Bank planning a new mechanism to fund strategic investments in homegrown tech. Meanwhile, the bloc is in bilateral talks with Washington — and floated a G7 "trusted partner" scheme with President Donald Trump — to keep access to the most advanced models.

Read at FT ↗

II.China Watch

Chinese frontier lab Z.ai switches on a 1-gigawatt data center with no Nvidia chips inside

Z.ai, the Beijing frontier lab formerly known as Zhipu, has completed a roughly 1-gigawatt data center powered entirely by domestically made AI chips, with parts already running training for its GLM model series, per Pandaily and Bloomberg. The company also acquired Zhongke Jiahe, a heterogeneous computing software firm spun out of the Chinese Academy of Sciences' computing institute, whose compiler technology helps run the same model across a fragmented mix of Huawei, Cambricon and other domestic accelerators. Z.ai is in early talks with domestic chip designers about custom silicon of its own, per The Information. The buildout follows Z.ai's January 2025 placement on the U.S. Commerce Department's Entity List, which cut off Nvidia hardware. Its stack of data center, compiler and prospective chip design shows Chinese frontier labs building around U.S. export controls rather than waiting for them to loosen.

Read at Pandaily ↗ Read at Bloomberg ↗ Read at The Information ↗

Chinese GPU maker Moore Threads shows a 256-card domestic supernode at WAIC

Moore Threads, a Chinese GPU designer, unveiled the MTT C256 supernode at the World Artificial Intelligence Conference in Shanghai, TechNode reported. The system packs 256 of its own GPUs into two racks and links them on a single network layer with under a microsecond of latency. The company demonstrated training of a 236 billion parameter mixture of experts model on more than 25 trillion tokens on the system. Multiple other Chinese GPU vendors, including Huawei, Biren and Muxi, showcased similar supernode products at WAIC, in an industry shift Chinese analysts have dubbed 2026's first year of domestic supernode production, per Pandaily. The demo lands as the Ministry of Industry and Information Technology anchors an emerging domestic standard through its CLink interconnection protocol, which several of the WAIC supernodes supported. The push is pulling Chinese GPU vendors toward a common playbook of networking many midtier chips together rather than competing chip-for-chip.

Read at TechNode ↗ Read at PanDaily ↗

Chinese police detail how they used AI to break into and monitor the dark web

Chinese police scientists have published research describing how they used AI tools to infiltrate dark web sites and decode underground content, a rare official look at a domestic surveillance capability, the South China Morning Post reported. Traditional web crawlers had struggled with the dark web's shifting login requirements, Captcha checks and anti-scraping tactics, so the researchers built AI tools to collect data and classify sensitive material once inside. The dark web relies on encrypted communications and anonymous routing that render user activity nearly untraceable, making it a common venue for illegal trafficking and data sales. The disclosure adds another AI capability to the surveillance stack Chinese police have publicly credited for monitoring encrypted platforms including Telegram and VPN traffic, per prior SCMP. It marks a rare instance of Beijing's security services documenting AI tooling in the open rather than deploying it quietly.

Read at South China Morning Post ↗

III.Policy Tracker

Energy Department appliance efficiency rollback advances as data center load strains the grid

The Trump administration is undoing federal fuel- and electricity-saving standards even as war in Iran, data centers and heat waves push up gasoline and electricity costs, the New York Times reported. The Energy Department earlier this month proposed to "Permanently End Green New Scam Appliance Mandates," taking aim at conservation rules for air conditioners, refrigerators and washing machines that Congress has directed to be continually tightened, per the Washington Post. If finalized after public comment, the rollback would prohibit further updates and could add gigawatts to residential demand. At least 1,662 Energy Department webpages offering guidance on protecting the electrical grid during heat waves went dark as of July 3, per The Guardian.

Read at NYT ↗ Read at Washington Post ↗ Read at The Guardian ↗

IV.Capability & Research Watch

Hugging Face discloses network breach by autonomous AI agent framework

Open-source AI platform Hugging Face said an autonomous AI agent framework breached its production infrastructure, accessing internal datasets and credentials, per BleepingComputer. Attackers seeded a malicious dataset that triggered two code-execution vulnerabilities on a processing worker, then stole cloud and cluster credentials and moved laterally across several internal clusters, the company said in an incident disclosure. The framework ran "many thousands of individual actions across a swarm of short-lived sandboxes, with self-migrating command-and-control staged on public services," Hugging Face said. The company said its software supply chain has been "verified clean" and it has found no evidence of tampering with public models, datasets or Spaces; the platform hosts more than 45,000 models used by more than 50,000 organizations.

Read at BleepingComputer ↗ Read at Hugging Face ↗

Sandbox escapes hit Cursor, OpenAI Codex, Google Gemini CLI and Antigravity

Pillar Security researchers broke out of the sandboxes on four widely used AI coding agents without attacking the sandbox itself, publishing a five-day "Week of Sandbox Escapes" series that includes Cursor, OpenAI's Codex, Google's Gemini CLI and Antigravity, per BleepingComputer. The agents obeyed sandbox rules but wrote files inside the project workspace; trusted tools outside the sandbox, such as VS Code task runners, Git integrations and Docker sockets, then read and executed them. That turns any prompt-injected instruction the agent writes into a command on the host machine, sidestepping the isolation model most vendors advertise.

Read at BleepingComputer ↗

Princeton and Chicago researchers find LLMs form hiring biases on the job, not just from training data

Researchers at Princeton and the University of Chicago ran ChatGPT, Claude and Gemini through a simulated 40-round hiring game and found the models quickly began stereotyping candidates from four fictional ethnic groups even though all candidates were equally likely to succeed, per MIT Technology Review. Each model was told it was a mayoral consultant hiring for 20 jobs, and after learning outcomes it started segregating groups into different roles. The finding suggests LLMs develop biases from experience, on top of biases picked up from training data, and that agentic memory features designed to retain user-specific context may accelerate the effect. The research was posted this month on OpenReview.

Read at MIT Technology Review ↗ Read at OpenReview ↗

Study finds ChatGPT and Gemini give inaccurate, volatile voting advice in Hungarian election test

General-purpose AI chatbots produced inaccurate and inconsistent voting guidance in a Hungarian election field study by the civil liberties group Liberties, misclassifying voter profiles, omitting relevant parties and listing parties not on the ballot, per The Guardian. Identical prompts drew "materially different" answers on different runs. In 90% of cases when ChatGPT was fed a detailed Tisza-aligned profile, it failed to recommend the party that won the April election in a landslide over Viktor Orbán's Fidesz. The researchers tested the two most popular chatbots in Hungary, ChatGPT and Gemini, against party positions from the Voksmonitor voting advice app; nearly 30% of Hungarians identified as AI users.

Read at The Guardian ↗ Read at libertiesEU ↗

V.Industry & Market Watch

Oracle could face $7 billion collateral bill for Wisconsin data center project

Oracle could be required to post as much as $7 billion in collateral for its Wisconsin data center project under state utility rules, the Financial Times reported. We Energies and Oracle have asked Wisconsin regulators to loosen the financial protections tied to the project, per Wisconsin Public Radio. The two companies are seeking to shift more of the risk of stranded generation, transmission and pipeline assets off ratepayers if the data center anchor tenant leaves.

Read at FT ↗ Read at WPR ↗

TSMC in talks to raise chipmaking prices by up to 10% in 2027

Taiwan Semiconductor Manufacturing Co. has held talks with clients about raising prices by 5% to 10% starting in 2027 to cover the rising cost of manufacturing materials, equipment and overseas plants, per Nikkei Asia and Bloomberg. Mature-node production, covering 12nm, 16nm and 28nm technologies, would face increases of up to 10%. Negotiations began in June and concluded in July, with new pricing set to take effect at the start of 2027. TSMC does not comment on pricing, a company spokesperson told Reuters, but chief executive C.C. Wei said in June that he would like to hike prices while avoiding the abrupt increases some memory firms have imposed. TSMC is the primary contract manufacturer for Nvidia, Apple and AMD.

Read at Bloomberg ↗ Read at The Star ↗

Anduril unveils autonomous helicopter wingman drone with Archer at Farnborough

Anduril and aviation firm Archer unveiled Thunder, a fully autonomous armed reconnaissance tiltrotor drone designed to fly alongside AH-64 Apache and future MV-75 Cheyenne helicopters, at the Farnborough Airshow, per Breaking Defense. Anduril president Chris Brose said Thunder has been in development for about three years and draws on lessons from the company's YFQ-44A Fury drone, which the Air Force is producing under its Collaborative Combat Aircraft program. The drone is designed to fly deep into hostile airspace at low altitude, hunt down fleeting ground targets, and provide close-in defense against air-to-air drone attacks on friendly rotorcraft, per The War Zone. Anduril said the design also draws on lessons U.S. and European allies have learned in Ukraine.

Read at Breaking Defense ↗ Read at The War Zone ↗

Sony Music sues Udio over 30,000 songs it says the AI generator infringed

Sony Music Entertainment filed a new lawsuit in New York on Monday accusing AI music generator Udio of infringing the copyright of more than 30,000 of its songs, from Elvis Presley's "Hound Dog" to Beyoncé's "Say My Name," per The Verge. Sony said the list represents only a small portion of works Udio allegedly infringed. Sony, Universal and Warner had sued Udio and rival AI music startup Suno in 2024 for copyright infringement.

Read at The Verge ↗

VI.Global & Geopolitics

UK PM Burnham names Narayan first AI minister to attend Cabinet as he abolishes the science department

Andy Burnham on Monday named Kanishka Narayan the UK's minister for AI and gave him a seat in Cabinet, a first for the AI portfolio, per Bloomberg. In the same reshuffle, Burnham dissolved the Department for Science, Innovation and Technology and merged it into a new Department for Business, Innovation, Science and Trade under Jonathan Reynolds. Narayan will retain his existing brief covering online safety alongside AI.

Read at Bloomberg ↗ Read at Sifted ↗

South Korea plans free homegrown AI chatbot this year to cut reliance on OpenAI

South Korea will launch a free general-purpose AI chatbot powered by domestic technology by year-end and follow it with a more capable AI agent, Science Minister Bae Kyung-hoon said Monday, per The Star. Roughly half of South Korea's 51 million people used generative AI in February, and nearly 23 million of them used ChatGPT, per market research firm Wiseapp-Retail. "To persuade users to move away from foreign AI services they already pay for, we need to provide services that are even better," Bae told reporters. The launch is the first deliverable of Seoul's "AI for All" scheme, which prioritizes integrating domestic AI agents into public services.

Read at The Star ↗

UK water industry warns government's AI data center plans lack a cooling water strategy

Water UK, the trade body for the country's water companies, told MPs in a written briefing that the government's water forecasts are "fatally flawed" and "explicitly exclude" data centers, and that the country will not have enough water for planned AI data center expansion, per The Guardian. The government's AI growth zones policy, AI for science strategy and AI opportunities action plan make no mention of water, and the Environment Agency's most recent national framework for water resources, published in June 2025, also excludes data centers. Slough already has the country's highest concentration of data centers, whose cooling towers, chillers and humidification systems consume large volumes of water directly and indirectly through electricity generation.

Read at The Guardian ↗