AI Policy · Daily

Beijing is weighing curbs on foreign access to advanced Chinese AI models, with the Commerce Ministry meeting Alibaba, ByteDance and Z.ai about limits that could extend to unreleased systems and criminalize leaks under national security law. House investigators are pressing Airbnb and Anysphere over their use of Chinese AI, seeking employee briefings on how the firms picked Alibaba's Qwen and Moonshot's Kimi. China's National Vulnerability Database flagged a serious "backdoor" risk in Anthropic's Claude Code, saying versions 2.1.91 through 2.1.196 can transmit user location and identity to remote servers without consent and advising organizations to uninstall or upgrade. The U.S. semiconductor workforce is projected to fall up to 157,000 workers short by 2030, with a McKinsey, SEMI and National Science Foundation analysis warning the gap threatens billions of dollars in new fab construction.

I.Top Stories

Beijing weighs restrictions on overseas access to top Chinese AI models

China's Ministry of Commerce has held meetings over the past month with Alibaba, ByteDance and startup Z.ai to discuss restricting overseas access to the country's most advanced AI models, including systems not yet released, three people familiar with the discussions told Reuters. The talks covered both closed source and open weight models, and considered making the leak or theft of proprietary AI technology an offense under China's national security law. Chinese officials were particularly concerned that Anthropic's Mythos model could be used to identify software vulnerabilities against Chinese interests, per two of the sources. The scope remains under discussion and may apply only to future models.

Read at Reuters ↗ • Read at Business Standard ↗

House committee probes Airbnb, Anysphere over adoption of Chinese built AI models

An ongoing House committee investigation is probing risks from U.S. firms adopting AI models built in China, CNBC reported Wednesday. The probe stems from April letters signed by House Homeland Security Chairman Andrew Garbarino, R-N.Y., and China Select Committee Chairman John Moolenaar, R-Mich., to Airbnb and Anysphere, per Nextgov/FCW. The letters requested details on the companies' use of Alibaba's Qwen and Moonshot AI's Kimi, and asked employees involved in model selection decisions to participate in an in-person briefing with lawmakers. Anysphere's Composer 2 coding model is built on Kimi, developed by Beijing-based Moonshot AI.

Read at CNBC ↗ • Read at Nextgov/FCW ↗

China's industry ministry issues 'backdoor' alert on Anthropic's Claude Code, warns versions 2.1.91 to 2.1.196

China's National Vulnerability Database, a cybersecurity platform run by the industry ministry, said Wednesday it had identified a serious "backdoor" risk in Anthropic's Claude Code, per Reuters. The NVDB said Claude Code contains a built-in monitoring mechanism capable of transmitting user location and identity to remote servers without consent, and applied the warning to versions 2.1.91 through 2.1.196. NVDB advised organizations to uninstall affected versions or upgrade to a release in which the alleged backdoor code has been removed, and to tighten controls on external network access for development tools. China's Alibaba banned employees from using Claude Code at work, Reuters reported last week.

Read at WSJ ↗ • Read at WHBL ↗

Report projects U.S. semiconductor workforce will fall short by 157,000 workers by 2030, delaying new fab construction

A new analysis from McKinsey, SEMI and the National Science Foundation projects that the U.S. semiconductor workforce will fall short by up to 157,000 full-time-equivalent workers by 2030, threatening construction of billions of dollars in new chip plants, per Bloomberg. The shortfall is expected to be deepest in Texas, California, Arizona, New York and Ohio, where new fab construction is concentrated. Nearly three-quarters of semiconductor employers surveyed already report difficulty recruiting engineers. About 3% of U.S. engineering students go into semiconductors, with most opting for higher paid software fields including AI.

Read at Bloomberg ↗

Americans nearly three times as likely to be concerned as excited about AI, new biweekly tracker finds

Americans are nearly three times as likely to be concerned as excited about AI's growing role in society, with 28% very concerned and 37% somewhat concerned versus 6% very excited and 18% somewhat excited, according to the debut "Americans on AI" poll from the new nonprofit Athena Insights, first shared with The Hill. About 70% said AI is coming into their lives whether they want it or not, and a similar share said they struggle to tell what is real because of AI. By nearly 20 percentage points, more Americans said the government is not doing enough about AI's effects on children and young people, the environment, and workers and jobs. The poll of 1,814 Americans, fielded June 25-30 at NORC at the University of Chicago with a margin of error of 3.4 percentage points, will run every other week for the next year.

Read at The Hill ↗

II.China Watch

DeepSeek is quietly designing its own AI inference chip to cut Nvidia reliance

DeepSeek has spent about a year building an in-house inference chip, per Reuters. The Hangzhou lab has held talks with chip design firms, wafer foundries and memory suppliers, and is recruiting chip design engineers without posting to public job boards. The chip targets inference rather than training, aligning with DeepSeek's reliance on Huawei's Ascend line and its record of extracting performance from constrained hardware. OpenAI unveiled its first custom inference chip with Broadcom this year, part of a broader shift as top model labs move to control their own silicon.

Read at Reuters ↗ • Read at QbitAI ↗

Chinese startup unveils a real-time interactive world model built entirely on Huawei chips

Moxin Technology, working with a Zhejiang University team led by academician Pan Yunhe, unveiled MoWorld, a 14 billion parameter "Flash World Model," per 36Kr. World models generate interactive 3D environments used to train robots and driving systems, and had been stuck at 5 to 10 frames per second on GPU stacks, the report said. The team said MoWorld runs above 50 frames per second on Huawei's Ascend 910C platform, with training, distillation and inference all on domestic chips and inference costs claimed at 70% below a comparable Nvidia GPU setup. The startup recently closed a round of more than $100 million backed by state-linked capital, a Middle East fund and more than 10 industrial investors, following earlier investment from Huawei's Hubble Investment.

Read at 36Kr ↗

China's brain-computer interface sector clears a wave of regulatory and funding milestones

NeuroXess started clinical trials for a subdural brain-computer interface implant at Shanghai's Huashan Hospital on Tuesday, Pandaily reported. Boruikang, whose device became China's first approved invasive BCI medical implant earlier this year, is pursuing an IPO on Shanghai's STAR Market, its tech exchange. First-quarter 2026 venture funding into the sector already exceeded all of 2025, with Alibaba leading a round at Ladder Medical and additional raises at BrainCo and Zhiran Medical. Shanghai's Huihuibao supplemental commercial health insurance now covers BCI surgery consumables, and Guangdong province added BCI to its medical pricing catalog. Both moves align with China's 15th Five-Year Plan, its 2026-2030 economic blueprint, which designates BCI as a priority future industry.

Read at Pandaily ↗ • Read at Shanghai Municipal Government ↗

Tencent Hunyuan reportedly hires away another OpenAI multimodal researcher

Pandaily said Tencent's Hunyuan AI group is set to hire OpenAI research scientist Tian Yonglong as its new multimodal team lead, reporting to chief AI scientist Yao Shunyu. Yao himself left OpenAI for Tencent in September 2025, and Tian would follow the same trajectory: Tsinghua undergraduate, CUHK master's, MIT PhD, and stints at Google DeepMind and OpenAI, per the report. Tencent's multimodal stack is seen as trailing ByteDance's Doubao and Alibaba's Qwen, and the hire is framed as an attempt to close that gap with senior overseas research talent.

Read at Pandaily ↗

III.Policy Tracker

Commerce clears OpenAI for broad GPT 5.6 launch after government testing

The Commerce Department gave OpenAI the green light for a broad launch of its GPT 5.6 model, a source familiar with the situation confirmed to Axios Tuesday, and OpenAI expects a wide release this week. Testing was done by the Center for AI Standards and Innovation within Commerce, with OpenAI technical experts staying in Washington to address questions. OpenAI had been limited last month to a staggered release to government-approved entities, echoing restrictions placed on Anthropic's Mythos and Fable models. The clearance comes as AI firms and the government operate ahead of the concrete model release standards called for in President Trump's latest AI executive order, which have yet to be finalized.

Read at Axios ↗

Federal CIO Greg Barbaccia sets Aug. 31 exit, also vacating federal chief AI officer role

Greg Barbaccia, the U.S. federal chief information officer who also serves as federal chief AI officer, told the CIO Council in a Tuesday email that his last day will be Aug. 31, per FedScoop. Deputy Thomas Flagg, previously the Department of Education CIO, would be the default acting official on Barbaccia's departure. Barbaccia joined the Trump administration in January 2025 as its top IT official and has coordinated AI policy, oversight and adoption across federal agencies. Flagg was brought on as Barbaccia's deputy only months ago.

Read at FedScoop ↗

IV.Capability & Research Watch

Noma Security's 'GitLost' flaw let unauthenticated attackers siphon private GitHub repositories via AI workflows

Researchers at Noma Security disclosed a critical prompt injection vulnerability in GitHub's new Agentic Workflows feature, per SiliconANGLE. The flaw let an unauthenticated attacker siphon data from private code repositories by posting a single crafted issue in a public one. The workflows run on an AI agent backed by either Anthropic's Claude or GitHub Copilot, and compile down to GitHub Actions, the platform's automated task system.

Read at SiliconANGLE ↗

AI models are outgrowing the benchmarks that measure their hacking skills

Federal agencies have until Aug. 1 to establish a classified benchmarking process for assessing frontier AI models' capabilities, and those standards may arrive as soon as this week, per Axios. Testing lab Irregular, which works with OpenAI, Anthropic and governments, released a cyber benchmark in late June measuring whether models can carry out offensive tasks such as remote code execution and privilege escalation. Armadin co-founder David Slater said his AI red teaming company's agents surpassed every public cyber benchmark within four weeks, and called the public tests "totally saturated" and "useless." Anthropic said last week it is building a standardized benchmark with Amazon, Google, Microsoft and other partners that focuses on the outcomes of a jailbreak rather than whether one is possible.

Read at Axios ↗

'HalluSquatting' pull-based prompt injection exposes nine popular AI tools to botnet assembly

Researchers documented "HalluSquatting," a new pull-based prompt injection technique that exposes nine popular AI coding assistants and agents to weaponization for large-scale attacks, per Ars Technica. Unlike push attacks, which must be sent to each target one at a time, HalluSquatting exploits large language models' tendency to hallucinate the locations of code repositories and agent skills rather than say "I don't know." Attackers can register the predictably hallucinated names, seed them with malicious code and infect machines at scale.

Read at Ars Technica ↗

V.Industry & Market Watch

Skilled labor shortage now testing U.S. data center construction as hyperscaler order books swell

Craft labor providers who build U.S. data centers are hitting growth limits after several quarters of record order books, per Bloomberg. Electricians, high-voltage technicians, HVAC specialists and other technical trades are among the hardest positions to fill, per Tom's Hardware. Associated Builders and Contractors estimates the industry needs to attract 349,000 net new workers in 2026 and 456,000 in 2027 to meet demand, per Insurance Business. Meta earlier this year partnered with CBRE, the commercial real estate firm, on a training initiative to expand the qualified worker pipeline for data center construction and operations.

Read at Bloomberg ↗ • Read at Tom's Hardware ↗ • Read at Insurance Business ↗

Meta rolls out Muse Image; Instagram users must opt out to block AI photo use

Meta on Tuesday launched Muse Image, the first image generation model built by its Superintelligence Labs division, and made it the engine for image tools across the Meta AI app, Instagram and WhatsApp, per Wired and The Verge. Instagram users with public accounts must opt out to block their photos from appearing in AI generated images created by other users. Alexandr Wang, who leads Meta's Superintelligence Labs, said on Threads that Muse Image is "agentic," working with the Muse Spark large language model "to reason through your prompt, search the web, and plan before it generates."

Read at Wired ↗ • Read at The Verge ↗ • Read at Meta ↗

VI.Global & Geopolitics

South Korea's punitive damages 'fake news' law takes effect, covers YouTube creators and large accounts

South Korea's new law allowing courts to award punitive damages of up to five times proven losses against news outlets and large social accounts that spread false or manipulated content took effect. Distributing content more than twice after a court has confirmed it false or manipulated can trigger a media regulator fine of up to 1 billion won (about $656,000). Platforms with more than 1 million daily users must pull flagged content or suspend accounts once someone reports it, with platforms making the initial call on whether content qualifies. The measure applies to YouTube creators alongside traditional news outlets and other large account operators.

Read at AP ↗ • Read at The Next Web ↗

Australia's AI Safety Institute begins testing latest frontier models as junior minister warns of deceptive behavior

Australia's assistant minister for technology Andrew Charlton said in a Sydney speech Tuesday that AI models are "cheating, deceiving and going their own way," as the federal AI Safety Institute begins testing the latest frontier models, per The Guardian. Charlton said "AI systems are already doing things their creators never intended" and that the time to get ahead of it is during testing. The government also launched two research projects alongside the model testing program.

Read at The Guardian ↗

Londoners fight Brick Lane data center plan, arguing housing needs outrank City compute demand

Residents and Tower Hamlets council are opposing a planned data center in Brick Lane in east London, arguing affordable housing needs are more urgent than added compute capacity for "high-frequency trading" in the nearby City, per The Guardian. Campaigners said the project would worsen the area's housing crisis and drive long-term residents away. The Brick Lane fight is part of the rapid UK data center rollout aimed at meeting demand created by AI.

Read at The Guardian ↗

ILO finds 80 million ASEAN workers exposed to generative AI, but no mass disruption yet

The International Labour Organization estimated in a new report that 22.9% of total employment across ASEAN's 11 Southeast Asian nations, or nearly 80 million workers, is in occupations with more than minimal exposure to generative AI, per MirageNews. Only a small share of workers are in the highest exposure occupations. The ILO said it found no evidence of large-scale job losses to date. The report is titled "Generative AI and labour markets in ASEAN: Significant exposure, limited disruption, uneven preparedness."

Read at MirageNews ↗ • Read at International Labour Organization ↗

EU opens data sovereignty consultation under Tech Sovereignty Package, comments close Sept. 8

The European Commission on Wednesday opened a targeted consultation on safeguarding EU data sovereignty, seeking views on data dependencies affecting European organizations, barriers to accessing or using data in third countries, obstacles to transferring data to the EU, and risks of third country access to sensitive data. The consultation is open to actors across the data value chain and closes at 23:59 CEST on Sept. 8, 2026, per the EU Digital Strategy notice. It follows the November 2025 Data Union Strategy and links to the European Tech Sovereignty Package covering semiconductors, AI, cloud and open source.

Read at EU Digital Strategy ↗