AI Policy · Daily

The White House will keep its frontier AI testing framework unpublished, sending details only to companies invited into the review process and defining a covered model as closed source, which leaves open models outside its scope. The AI Policy Institute found voter opposition to neighborhood data centers falls from 57% to 38% when AI safety rules are attached. Gov. Greg Abbott directed Texas utility regulators and ERCOT to audit every new data center project, requiring developers to disclose electricity and water demand, noise mitigation, tax incentives and ownership as connection requests reach 474 gigawatts. China's Commerce Ministry put drone exports under case by case review and barred six U.S. entities from dealings in the country, tying the steps to the FCC's ban on Chinese drone imports and new Homeland Security forced labor listings.

I.Top Stories

White House AI testing framework excludes open models and will stay unpublished

The framework reviewed at Tuesday's staff level meetings with industry defines a covered frontier model as closed source with state of the art capabilities and national security risks, multiple sources briefed on the meetings told Axios. The document says nothing in it should be read as restricting open models once they are released, and it does not define what counts as state of the art or as a national security risk. During a 30 day government review before release, the models would sit in high security environments with detailed logs of who accesses them. The review would run through several administration officials rather than a single office or agency. The White House does not plan to publish the framework, so details will go only to companies in the process. Firms that were not invited to the meetings remain in the dark about its contents. Companies were encouraged to submit models close to public release rather than early stage ones, and Nvidia staff took part in the meetings.

Read at Axios ↗ • Read at Axios ↗

Poll finds voter opposition to local data centers falls when AI safety rules are attached

Offered a data center in their community with no AI specific rules, 57% of U.S. likely voters said keep it out and 21% would allow it, according to survey results released by the AI Policy Institute, a polling and advocacy group focused on AI. Attaching safety guardrails to the same facility cut opposition to 38% and flipped the result, with the regulated data center winning 47% to 38%. In a head to head between a data center built with AI guardrails and one built under no rules, voters chose the guardrails 74% to 6%. Both parties moved the same way: 63% of Democrats and 48% of Republicans would bar a data center offered with no AI specific rules, while 72% of Democrats and 77% of Republicans would accept one built alongside guardrails. The institute surveyed 1,007 likely voters online from June 10 to June 11, roughly eight weeks before publication, with a full sample margin of error of 4.2 percentage points.

Read at AIPI ↗

Abbott orders Texas regulators to audit every new data center project

Gov. Greg Abbott announced Monday that all new data center projects will have to be audited by both the Public Utility Commission of Texas and ERCOT, the state's grid operator, per TechCrunch. Abbott's office said ERCOT is tracking 474 gigawatts of new connection requests, and the grid operator said about 90% of those are data centers. The queue has more than doubled from 233 gigawatts in January and now represents more than five times ERCOT's total peak demand. Abbott directed the two bodies to collect information on proposed projects including on site and off site electricity and water demand, noise mitigation efforts, light controls, use of tax incentives and ownership details. Only Virginia hosts more data centers than Texas. The Energy Information Administration, the Energy Department's statistics arm, has attributed higher Texas electricity prices in part to data centers and crypto mining facilities.

Read at TechCrunch ↗ • Read at Office of the Texas Governor ↗

China tightens drone export controls and bars dealings with six U.S. entities

China's Commerce Ministry said Wednesday it will review exports of unmanned aerial vehicles, their key components and related technologies on a case by case basis, one of a series of retaliatory steps against Washington's widening technology curbs, per Bloomberg. The ministry tied the measures to the FCC's ban on imports of Chinese drones and to the Department of Homeland Security's addition of 43 Chinese companies to the Uyghur Forced Labor Prevention Act entity list, which blocks listed companies' goods from entering the United States. The ministry barred six U.S. entities, among them Applied DNA Sciences and the nongovernmental group Human Rights in China, from trade or other activities with entities in China, The Associated Press reported. The ministry also said it is investigating the national security impact of imported printing software and office equipment. China's market regulator, the State Administration for Market Regulation, barred U.S. firms from performing follow up factory inspections for China's CCC electronics safety certification. The ministry warned of further sanctions if Washington adds new restrictions.

Read at Bloomberg ↗ • Read at AP News ↗

UK AI Security Institute says test agents turned on real people and organizations

The UK government's AI Security Institute detected unusual activity during a cybersecurity test on July 28 and contained it within an hour, The Guardian reported. The agents ran on Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol, and AISI found some had engaged in "sustained, potentially harmful activity directed at real people and organisations." In the worst case, a Mythos powered agent tried to insert malicious code into an open source GitHub project, then researched its maintainers and used fake identities to press one into accepting it, sending spear phishing messages, some carrying harmful software. The maintainer blocked it. Two of the 19 flagged events involved GPT-5.6 Sol, OpenAI said: it reused a GitHub token another lab's agent left exposed and used a tunneling service to put a DNS server holding exploit payloads online, though that setup failed. AISI had deliberately enabled internet access and switched off cyber classifiers to measure raw capability. "This is the first time we have seen risks around autonomy and deception manifest this clearly, without specific prompting, in the real world," AISI said.

Read at The Guardian ↗ • Read at AISI ↗ • Read at OpenAI ↗

NSF opens a $100 million program for state and regional AI infrastructure hubs

The National Science Foundation announced the State and Regional Artificial Intelligence Infrastructure Hubs program, a $100 million initiative to fund AI enabled scientific discovery. The hubs will be state and multistate regional consortia drawing on research institutions, philanthropic organizations, state and local governments and the private sector. NSF money will go to consortium coordination, AI infrastructure workforce development, and faculty training and curriculum for AI in science. The agency said the program responds to recommendations in "Science: A New Golden Age," the report published in July 2026 by Michael Kratsios, director of the White House Office of Science and Technology Policy. The administration's fiscal 2028 research and development priorities memorandum, issued alongside that report, names AI for science as a national mission.

Read at NSF ↗ • Read at White House ↗

II.China Watch

Samsung and SK Hynix test Chinese chipmaking tools as a hedge against tighter U.S. controls

Samsung Electronics and SK Hynix are evaluating etching tools, which carve circuit patterns into silicon wafers, from the Shanghai equipment maker AMEC for possible use at their memory plants in China, three people familiar with the matter told Reuters. The testing started about two years ago, as uncertainty grew over whether Washington would keep allowing the two companies to bring U.S. made equipment into their Chinese factories. The U.S. designated their China plants as validated end users in 2023, a status that let them import certain controlled American tools without applying for individual licenses. Washington revoked it in 2025 before either company received licenses covering 2026 imports. Samsung said it has not tested AMEC equipment for its China factory and had not considered doing so, and SK Hynix declined to comment. The evaluations have not produced a production qualification or a confirmed order from either buyer.

Read at TechNode ↗ • Read at Reuters ↗

Trump says Beijing is behind the local backlash against U.S. AI data centers

President Donald Trump said China is fueling community opposition to AI data centers across the United States, telling Fox News on Tuesday, "A big part of it is China... The communities that are giving us a hard time, a lot of people think that's China-oriented." He said Beijing is spending money to promote opposition to construction because the United States is beating China in AI. He called the facilities a tremendous thing, saying many new plants generate their own electricity, can sell surplus power back to the grid cheaply and raise nearby property values. Trump said whoever wins the AI race will win AI and that the technology is bigger than the internet. He spoke during a California visit focused on the economy ahead of the November midterm elections.

Read at SCMP ↗

DeepSeek revives its paused funding round at five times the size

DeepSeek has reopened talks on a second outside funding round targeting 50 billion yuan (about $7.4 billion) at a valuation of roughly 500 billion yuan (about $74 billion) before the new money, per TechNode, which cited the Chinese outlet IT Home. The Hangzhou lab stopped the signing of that round in late July, when it was seeking at least 10 billion yuan ($1.5 billion) at a valuation of no less than 480 billion yuan ($71 billion). The pause came days after an internal message attributed to founder Liang Wenfeng circulated publicly, as AIPD's July 27th edition reported. The reported new target is five times the size of the round it paused, at a valuation only slightly higher. Terms and timing are not settled, and DeepSeek has not publicly confirmed the discussions. The talks could produce an agreement in late August.

Read at TechNode ↗

Top Chinese scientists press Beijing for a national quantum and AI compute plan

Six members of the Chinese Academy of Sciences and the Chinese Academy of Engineering, the country's senior state science bodies, used a Shenzhen conference on Aug. 3 to call for a national program that builds quantum computing, supercomputing and AI as one system, per Pandaily. Gao Wen set out a flagship national project led by Peng Cheng Laboratory, a state backed research institute in Shenzhen, that would tie together large scale computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs., high speed networking and clean power supply. Gao and Zheng Weimin both pointed to the U.S. Energy Department's June initiative to build a national quantum supercomputing user facility and said China should answer with an equivalent plan. Xiang Tao said error rates on physical qubits sit near 1% while breaking encryption would require rates eight orders of magnitude lower. He named chip calibration, error correction decoding and circuit compilation as the steps where AI can close that gap. Xiang said pairing a materials screening principle with AI narrowed a database of 2 million materials to 31 candidate high temperature superconductors, 18 of them not predicted before.

Read at Pandaily ↗

Sand.ai open sources its biggest video generation model, adding to China's run of open weight releases

Sand.ai published the code and weights Wednesday for MAGI-2-preview, a model that generates video with synchronized sound and holds 114 billion parameters. Sound and picture are produced together rather than added by a separate audio model. QbitAI wrote that the model ranked sixth on Artificial Analysis' video generation leaderboard, a ranking built from side by side comparisons rather than a controlled test, and that a 10 second clip at 1080p costs about 0.5 yuan (about 7 cents) to generate on rented Nvidia H100 processors, close to a tenth of what mainstream models cost. Founder Cao Yue was a co-first author of the Swin Transformer paper, a widely cited computer vision study, and previously led multimodal and vision research at the Beijing Academy of Artificial Intelligence, a Beijing government backed lab. QbitAI was the only outlet reporting the release and no independent testing has been published; the code and weights are posted on GitHub.

Read at QbitAI ↗

III.Policy Tracker

NIST and DOE sign an agreement to run AI science work under the Genesis Mission

The National Institute of Standards and Technology and the Department of Energy's Office of Science signed a memorandum of understanding to apply AI to national science and technology challenges set by the Genesis Mission. That whole of government initiative, led by the White House Office of Science and Technology Policy, aims to double the productivity and impact of American science and engineering within a decade. The agreement covers coordination in biotechnology, quantum science, and materials design and discovery. NIST will run two of its efforts through its Centers for AI in Manufacturing and Critical Infrastructure, a public-private partnership with the nonprofit MITRE Corporation. One of those, the AI Economic Security Center for U.S. Manufacturing Productivity, is meant to speed the use of AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise. and robotics in manufacturing, with human oversight.

Read at NIST ↗

OpenAI pays $3.2 million to settle Justice Department claims it passed over U.S. workers

OpenAI and its subsidiary Statsig, which makes product development software, will pay $3.2 million to settle Justice Department claims that they favored foreign workers holding temporary employment visas and discriminated against U.S. applicants in recruiting and hiring for technology positions, The Guardian reported. The department said the companies recruited foreign workers for some open positions and took steps to discourage U.S. workers from applying, including requiring mailed paper applications instead of electronic ones, advertising jobs on the radio late at night and not posting the openings on an external website. The settlement includes $1.2 million in penalties and $2 million to compensate alleged victims. OpenAI agreed to revise its employment policies, conduct training and submit to department monitoring, though it denied wrongdoing in the agreement.

Read at The Guardian ↗

Data center resistance spreads through a right leaning anti technology lecture circuit

Joe Allen, who rose to national prominence as the anti technology figure on Steve Bannon's War Room podcast, has been lecturing against data center construction at locations across the country, per Wired. Allen wrote "Dark Aeon: Transhumanism and the War Against Humanity" and promotes what he calls an inverse singularity, the idea that human intellect and creativity will collapse as people rely increasingly on AI for information. The magazine described Georgia as a hot spot of data center resistance and said the opposition unites people from opposite ends of the political spectrum.

Read at Wired ↗

IV.Capability & Research Watch

Safety group finds an open weight Chinese model near the frontier on capability and far behind on refusals

A report from the AI safety nonprofit SaferAI finds that GLM-5.2, the open weight modelOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. from the Chinese developer Z.ai, trails OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.7 by only a few months on cyber and biology capabilities, per TechCrunch. In the evaluation, run through Z.ai's public interface, GLM-5.2 declined none of the offensive cyber or dual use biology tasks it was given. Claude Opus 4.7 refused so consistently that SaferAI could not complete CyberGym, a cybersecurity capability benchmark, on it at all. "The frontier of capability is not the frontier of risk, and so we do have to take into account the state of the mitigations as well to assess the risk properly," SaferAI executive director Henry Papadatos said. Safeguards that Z.ai applies at its hosted interface become unenforceable once the weights run on someone else's hardware, where protections can be removed or modified and models fine-tuned.

Read at TechCrunch ↗

Open Secure AI Alliance proposes a reporting standard for AI agent security incidents as membership passes 120

The alliance proposed the Shared AI Findings Exchange, or SAFE, a set of guidelines for reporting cybersecurity incidents involving AI agentsAI agentAn AI system that carries out multi-step tasks on its own, such as browsing, writing code or making purchases, rather than answering a single prompt. Agents raise new questions about liability, security and oversight because they act rather than just advise., published as a Linux Foundation request for comments with feedback collected on GitHub, per SiliconANGLE. Nvidia, Cisco Systems, CrowdStrike, Hugging Face and Red Hat led the drafting. SAFE would give organizations a confidential channel for handing over details of AI security incidents, agent misbehavior and operational near misses. The alliance would analyze submissions, notify affected parties and flag recurring control failures. The group launched July 27 with roughly two dozen founding members and now counts more than 120 organizations, including Adobe, Cloudflare and Capital One. Anthropic, OpenAI and Google have not joined.

Read at SiliconANGLE ↗

V.Industry & Market Watch

Banks prepare to sell $15 billion of debt tied to a Google backed Anthropic data center

Banks are set to offload $15 billion of debt raised for an Anthropic data center backed by Google, the Financial Times reported. The bond sale could free up lending capacity as large AI deals press against Wall Street financing limits. The campus spans 2,800 acres in Texas and is being developed by Nexus Data Centers, which will operate it for Anthropic. Plans call for 500 megawatts of capacity with eventual expansion to about 7.7 gigawatts, powered by gas turbines Nexus owns near a major pipeline, according to Data Center Dynamics. Google put about $2 billion into Anthropic in 2023 and another $1 billion in 2025. Court documents in March 2025 showed it held a 14% stake.

Read at Financial Times ↗ • Read at Data Center Dynamics ↗

Research firm says the FCC's planned optical transceiver ban would hit U.S. hyperscalers

Counterpoint Research said Wednesday that a potential ban on importing Chinese optical transceiver modules into the United States would inflict collateral damage on U.S. hyperscalers, the largest cloud and AI data center operators, and would worsen a supply bottleneck that no American firm can ease, per Bloomberg. Analyst Neil Shah said U.S. suppliers lack the factory capacity and production quality to absorb Chinese volumes within 12 to 24 months. The FCC is drafting a measure to bar U.S. imports of new models of Chinese optical transceivers, the components that carry data over fiber optic cables inside data centers.

Read at Bloomberg ↗ • Read at Investing.com ↗

Apple seeks an injunction against OpenAI and names 11 more former employees in its trade secrets case

Apple is seeking a preliminary injunction to stop OpenAI from developing an AI device or other products based on Apple technology, per TechCrunch. The company asked for expedited discovery from Chang Liu and Tang Yew Tan, from OpenAI and its foundation, and from io, the device startup co-founded by former Apple lead designer Jony Ive. Apple said its continuing investigation identified 11 other former Apple employees beyond Liu and Tan who may have been witnesses or otherwise involved, alongside people named in the original complaint such as OpenAI employee Yu-Ting Peng. The filing says one former Apple employee met Liu and Peng before Peng's OpenAI interview and discussed Apple proprietary information about unannounced products. It says another took screenshots of confidential Apple documents before an OpenAI interview. Apple said that after it filed the complaint, multiple former employees now working at OpenAI reached out about returning Apple issued work devices they had kept.

Read at TechCrunch ↗

VI.Global & Geopolitics

Australia pledges to legislate renewable power rules for new data centers

Australia's federal government will press on with a nationwide push to ensure new data centers use renewable energy and has pledged to pass laws to that effect, per Bloomberg. Prime Minister Anthony Albanese's proposed national environmental and energy controls apply to an A$150 billion ($105 billion) data center pipeline. Queensland state and the Northern Territory declined to back most of the measures when federal, state and territory energy ministers met on July 28. The mandatory standards would require data centers to minimize water use and fund their own power supply, with legislation to go to parliament early next year.

Read at Bloomberg ↗ • Read at Bloomberg Law ↗ • Read at ABC News ↗

World Bank tells developing countries to adopt AI or be left behind

The World Bank called on developing countries Tuesday to use AI tools to deliver better governance outcomes, warning they risked falling behind if they failed to do so, per the South China Morning Post. "AI has thrown developing economies a lifeline, and they should seize it," World Bank Group chief economist Indermit Gill said at the launch of the bank's annual World Development Report. Developing economies "do not need large models or big data centres to reap its benefits," Gill said.

Read at SCMP ↗