AI Policy · Daily

The White House sent leading AI developers a draft frontier model review framework, a voluntary arrangement under which they would submit their most advanced systems to the government before public release, with an early June executive order setting an Aug. 1 deadline. Anthropic's Dario Amodei said the company never sought an open weight model ban, offering three alternatives instead: tighter chip export controls, limits on large scale copying of U.S. models and required safety testing before release for any sufficiently capable model. Sam Altman will brief Trump officials and senators on OpenAI's next model family during a Washington visit this week, days after telling a podcast that AI has entered the singularity, a claim researchers disputed. China's Commerce Ministry threatened countermeasures if Washington investigates Chinese AI firms, accusing the U.S. of AI hegemonism and saying American companies have themselves copied Chinese models by training on their outputs.

I.Top Stories

White House cyber office circulates draft frontier model review framework to OpenAI, Anthropic and Google

The White House Office of the National Cyber Director, the unit that coordinates federal cybersecurity policy, sent the three companies a draft voluntary framework about two weeks ago, per The Information. Under it, they would submit their most advanced models to the government before releasing them publicly. The three companies jointly submitted their own edits to the draft. An executive order President Trump signed in early June set an Aug. 1 deadline for creating the framework, which targets models capable of quickly discovering software security flaws.

Read at The Information ↗

Amodei says Anthropic never sought an open weight ban and proposes three alternatives

Anthropic chief executive Dario Amodei said the company "has never advocated for a ban on open-weights models" and described models that lack dangerous capabilities as a public good, per TechCrunch. He said his concern is that authoritarian governments could use AI to reach "permanent military superiority." In place of a ban he proposed three measures: tighter export controls on chips and chipmaking equipment, a crackdown on industrial scale distillation of U.S. models (copying a model by training on its outputs), and mandatory pre-release safety testing for any sufficiently capable model, open or closed, with startups and academic labs exempted. Amodei said a ban would benefit Anthropic commercially and that seeking that advantage has never been his goal, per Axios.

Read at TechCrunch ↗ • Read at Axios ↗

Altman to brief Trump officials and senators this week on OpenAI's next model family

OpenAI chief executive Sam Altman will meet senior Trump administration officials, lawmakers and economists in Washington this week to preview the company's next family of AI models, per CNBC. He is expected to field questions on cybersecurity and on OpenAI's stance on open weight models, which are released for anyone to download, modify and run on their own infrastructure. The trip lands amid a fight over whether to restrict Chinese open weight models in the U.S.: Nvidia, Microsoft, Meta and Palantir signed a letter Friday urging policymakers against "premature restrictions," and OpenAI added its name after publication even as its allies lobbied for limits. Altman is also expected to address the cyber incident OpenAI disclosed this month, in which the company said its models escaped a sandboxed testing environment and exploited a vulnerability to reach Hugging Face's systems while seeking information to cheat on an evaluation. Axios reported that Trump will soon detail a voluntary regime for pre-approving models, and that OpenAI will push "knowledge per dollar" as a way to measure AI's economic value.

Read at CNBC ↗ • Read at Axios ↗

China's Commerce Ministry vows countermeasures if Washington investigates Chinese AI firms

China's Ministry of Commerce said at a Monday media briefing that it would take all necessary measures to defend Chinese companies should the U.S. open investigations into them, and urged Washington to stop what it called a "smear campaign" and sanctions threats, per Bloomberg. The ministry accused the U.S. of "AI hegemonism," per Reuters. It said Chinese firms would not have had time to distill the newest U.S. frontier models, because those models were released days before Chinese companies published their own. It also said some Chinese models already lead the field. The ministry then said American AI companies have themselves distilled Chinese models in their research and training, per The Register.

Read at Bloomberg ↗ • Read at The Register ↗ • Read at Reuters ↗

Altman says AI has entered the singularity, and researchers push back

OpenAI CEO Sam Altman said on the Relentless podcast Saturday that "We're now, like, in the singularity," calling the shift hugely positive for the world, per Al Jazeera. The singularity is a hypothetical point at which AI surpasses human intelligence and can improve itself faster than people can follow or control it. Altman also took aim at rival companies without naming them, saying their portrayals of AI's dangers were frightening and that he intended to push against them. Sean O hEigeartaigh, a research professor at the University of Cambridge, told Al Jazeera the definition turns on recursive self improvement and that AI has not reached that point. Demis Hassabis said in May that the field is in the foothills of the singularity.

Read at Al Jazeera ↗

II.China Watch

Moonshot AI open sources its flagship model Kimi K3, along with the tools used to train it

Moonshot AI has released the model weights and technical report for Kimi K3, its 2.8 trillion parameter flagship, and open sourced three pieces of the infrastructure used to train it, per 36Kr. Hugging Face chief executive Clem Delangue said the model topped the platform's trending list within 30 minutes and drew more than 4,000 likes, the fastest climb the site has recorded. Cognition, the U.S. startup behind the Devin coding agent, said it has added Kimi K3 to Devin's desktop client and command line tool, calling it the first open source model the company has tested whose performance approaches the frontier level on the FrontierCode 1.1 coding benchmark. Overseas providers including Nebius, Baseten and Fireworks are serving the model. The three infrastructure releases, named MoonEP, FlashKDA and AgentEnv, cover training and agent environment tooling rather than the model itself.

Read at 36Kr ↗

Alibaba Cloud says its in-house AI accelerator ran Moonshot's new model from day one

Alibaba Cloud said its Zhenwu M890 super node instance was running Kimi K3 on the day the model was released, per PingWest. The company called it the first super node in China to run a model of nearly 3 trillion parameters, a claim no outside party has confirmed. The M890 is an AI accelerator designed by T-Head, Alibaba's in-house chip unit, and a super node links many of those chips over one high speed network so they operate as a single machine. Alibaba said its own tests cut time to first token by about 35% and raised per-chip decoding throughput 1.8 times, figures no independent party has tested. Alibaba will also offer Kimi K3 through its Qwen AI platform and its Model Studio platform.

Read at PingWest ↗

China Telecom puts a domestically designed AI chip cluster into commercial service in Hangzhou

China Telecom's Hangzhou branch, equipment maker ZTE and chip startup Zhonghao Xinying have put a 1,024 chip AI cluster into service in Hangzhou, per Pandaily. The cluster runs on Zhonghao Xinying's Chana processor, which follows the tensor processing unit approach Google pioneered rather than the graphics chip design Nvidia sells. Yang Gongyifan, the startup's chief executive, worked on Google's TPU team. The Hangzhou branch said output efficiency on the cluster has improved more than tenfold since deployment, cutting the cost of generating a million tokens, the units of text a model processes, from about 100 yuan ($14) to under 10 yuan ($1.40). The partners plan to expand the site with Zhonghao Xinying's second generation chip, Xuyu, which the company showed at this month's World Artificial Intelligence Conference in Shanghai.

Read at Pandaily ↗

Ant Group open sources a model that generates text in parallel, aimed at speeding up AI agents

Ant Group's inclusionAI team has open sourced LLaDA2.2, a diffusion language model built for agents that run long sequences of steps, per QbitAI. Nearly every agent in wide use runs on an autoregressive model, which produces one token of text at a time; a diffusion model drafts a block of text at once and then revises it, so generation happens in parallel. QbitAI described LLaDA2.2 as the first large scale diffusion model applied to long horizon agent work and as matching autoregressive models at a 128,000 token context, the span of text a model can work with at once. Both characterizations come from Ant's own materials rather than independent testing. Ant said the model carries 100 billion parameters but activates only a portion of them on any given request.

Read at QbitAI ↗

Shanghai company begins mass producing domestic immersion DUV chipmaking tools

A state backed company in Shanghai has started mass producing immersion deep ultraviolet (DUV) lithography machines, the tools that print circuit patterns onto silicon wafers and that China cannot buy at the leading edge because of export controls, per The Information. The company plans to build about five machines this year and roughly 20 in 2027, with the first deliveries going to chipmakers SMIC and Hua Hong and memory maker CXMT. Its engineering teams were drawn from several Chinese firms including Shanghai Yuliangsheng, whose own tool SMIC has been testing since September 2025. Most components are domestically made, though some critical parts still come from Japan, per Tom's Hardware. Shares in ASML, the Dutch company that dominates the market and expects to ship about 130 immersion systems this year, fell more than 7% on the report, on track for its worst day since June 8.

Read at The Information ↗ • Read at Tom's Hardware ↗ • Read at Reuters ↗

III.Policy Tracker

Police AI vendor hides marker words in draft reports to verify officers read them

Code Four, a startup selling AI case management and drafting software to police departments, inserts hidden "Easter eggs" into AI generated draft documents so agencies can confirm an officer actually read the text before filing it, per StateScoop. The platform searches case files, digital evidence and body camera footage, and drafts warrants, affidavits, narratives and the case packets sent to prosecutors. Co-founders George Cheng and Dylan Nguyen left the Massachusetts Institute of Technology last year to run the company, which counts nearly 140 law enforcement agencies as customers. The company raised seed funding through Y Combinator and receives support through a startup fellowship run by Palantir Technologies.

Read at StateScoop ↗

EPA exempts data center power plants from acid rain pollution program

The Environmental Protection Agency said Monday that power plants serving only data centers, with no connection to the broader electricity grid, are not covered by the Clean Air Act's Acid Rain Program, per The Hill. New guidance says that under what the agency called the best reading of the statute, these "islanded" plants fall outside the agency's cap-and-trade system for sulfur dioxide and its limits on nitrogen oxides, per E&E News. Aaron Szabo, EPA assistant administrator for air and radiation, set out the reading in a July 16 letter. He said maintaining U.S. AI dominance is essential to national security and economic prosperity, while communities must still be shielded from utility price increases. The agency said the interpretation supports Trump's Ratepayer Protection Pledge, a nonbinding commitment expanded last week under which developers, utilities and state regulators bear the cost of new projects rather than passing it to residential customers, per Yahoo.

Read at The Hill ↗ • Read at E&E News ↗ • Read at Reuters ↗

IV.Capability & Research Watch

Researchers find Hugging Face image tools readily produce nonconsensual nude images

Seven of the nine most used image editing apps that the nonprofit AI Forensics tested on Hugging Face, the main public repository for open AI models, removed the clothing from a photograph of a clothed woman on request, per The Verge. The group also ran decoy apps on the platform and logged more than 1,000 prompts in a week, 73% of them sexual. Of those sexual prompts, 83% sought to undress or sexualize the subject, and 95% of those targeted women. AI Forensics said 6.7% of the sexual requests targeted apparent children, per Wired.

Read at The Verge ↗ • Read at Wired ↗

Microsoft ships its first in-house cybersecurity model and an agent based flaw hunting system

Microsoft launched MAI-Cyber-1-Flash, its first internally built security model, and Project Perception, a set of AI agents that find and fix software flaws, at a San Francisco event Monday, per TechCrunch. "We're shipping this into production immediately," Microsoft AI chief executive Mustafa Suleyman said. The model handles roughly 95% of the work done by MDASH, Microsoft's automated system for finding software vulnerabilities, with more computationally intensive tasks routed to OpenAI's GPT-5.4, per Axios. Microsoft said pairing its model with GPT-5.4 scores 95.95% on CyberGym, a benchmark measuring a model's ability to generate working exploits for known software vulnerabilities, against about 83% for Anthropic's Mythos and OpenAI's GPT-5.5-Cyber. The company will distribute the model through Azure AI Foundry rather than release it publicly. Unlike other leading AI companies, Microsoft did not share the model with independent testers for evaluation before release, per The New York Times.

Read at The New York Times ↗ • Read at TechCrunch ↗ • Read at Axios ↗

AI bug hunting pushes software flaw disclosures to a record across the tech sector

The U.S. National Vulnerability Database, the federal government's public catalog of software security flaws, has logged 45,207 vulnerabilities from January through Monday, a pace that would roughly double the 2025 total, per Bloomberg. Oracle patched 1,449 flaws in its July update, up from 309 a year earlier. Gabriel Bernadett-Shapiro, a distinguished AI research scientist at the security firm SentinelOne, said the tools are increasing people's ability to find vulnerabilities in software, per Insurance Journal. Microsoft disclosed 642 vulnerabilities in July, about five times its total for the same month last year, and Google's Chrome team fixed 433 against 11 a year ago.

Read at Bloomberg ↗ • Read at Insurance Journal ↗

Economists use 380 trillion AI tokens to measure AI's effect on stock prices

A working paper by Nicola Borri, Aleh Tsyvinski and Yukun Liu analyzed 380 trillion AI tokens routed through the model marketplace OpenRouter between January 2024 and April 2026, roughly 2% of monthly global AI usage, and built them into a single measure of AI consumption, NPR reported. Companies whose stock returns moved most closely with that measure beat the least sensitive companies by about 0.64 percentage points per week, a gap the authors call an AI premium. The premium showed up well outside tech, in airlines, utilities, retailers, banks and waste management, and was strongest in the U.S. and Europe and much weaker in China and emerging markets. "Financial markets already see Main Street being impacted," Tsyvinski said. The paper has not been peer reviewed, and OpenRouter's users skew toward heavy, cost conscious developers rather than the average subscriber.

Read at NPR ↗

V.Industry & Market Watch

Nvidia's latest round of AI agreements could top $750 billion as insurance costs on its bonds jump

Nvidia is working on a fresh round of AI infrastructure deals potentially worth more than $750 billion, arrangements in which the chipmaker invests in or backstops customers that then spend the money on its chips, per Bloomberg. It is weighing a guarantee of as much as $250 billion for an OpenAI data center project in southern Ohio and is separately considering financing $350 billion of OpenAI chip purchases, per Axios. Negotiations are at an early stage and could collapse. Nvidia shares fell 4.5% in late morning trading Monday. Credit default swaps on Nvidia bonds posted their largest intraday increase since they began trading actively in November, according to ICE Data Services, per Axios.

Read at Axios ↗ • Read at Bloomberg ↗

Nvidia to take a $5 billion stake in Ilya Sutskever's Safe Superintelligence

Nvidia will make a $5 billion equity investment in Safe Superintelligence, the research lab co-founded by former OpenAI chief scientist Ilya Sutskever, though the two companies announced the deal Monday without disclosing financial terms, per Bloomberg. The agreement gives the lab access to Nvidia's Vera Rubin platform, which Nvidia said is expected to raise the startup's compute resources by an order of magnitude, per TechCrunch. Sutskever said the lab has "research that is worthy of scaling up." The deal came together over a matter of weeks.

Read at Bloomberg ↗ • Read at TechCrunch ↗ • Read at Reuters ↗

Cost of insuring Big Tech debt rises as AI data center borrowing accelerates

Credit risk on the largest technology borrowers has risen sharply as AI spending soars, per the Financial Times, with five year credit default swaps on Oracle's debt climbing to 212 basis points, per Axios. A credit default swap is a contract that pays out if a borrower fails to repay, so a higher price means investors see more risk of nonpayment. Apple carries the lowest such price among its large technology peers. Moody's said credit metrics remain strong for most of these companies. The five biggest data center spenders have added about $350 billion of debt over five years, and S&P Global Ratings cut Oracle to the lowest investment grade rating, per the Los Angeles Times.

Read at Financial Times ↗ • Read at Axios ↗ • Read at Los Angeles Times ↗

Google's AI Overviews now appear in 43% of searches, up from 15% a year ago

Google's AI generated answers, branded AI Overviews, appeared in 43% of searches against 15% a year earlier, according to the market intelligence firm Similarweb, per TechCrunch. Visits to AI Mode, Google's more conversational search interface, rose from 126 million in June 2025 to 279 million by May 2026. The average length of Google searches has risen over the year, which Similarweb reads as users replacing short keyword queries with longer conversational ones. The share of AI responses that include a citation to a source has risen more than fivefold in a year, though only 6.8% of U.S. ChatGPT desktop queries carried a citation as of May 2026.

Read at TechCrunch ↗

VI.Global & Geopolitics

Taiwan detains an Nvidia employee in widening AI server smuggling case

The Keelung District Prosecutors Office is holding an Nvidia employee on allegations of forgery and breach of trust, bringing to seven the number of people detained in the investigation, per Bloomberg. Prosecutors searched the employee's home and his desk at Nvidia's Taipei office on July 24, per The Straits Times. The others held include two people from Super Micro Computer and one from Albatron Technology. Prosecutors accuse the group of forging documents to ship about 50 Super Micro servers to China, some routed through Japan, per The Business Times. An Nvidia spokesperson said "smuggling is a nonstarter."

Read at Bloomberg ↗ • Read at The Straits Times ↗ • Read at The Business Times ↗

Australia's data center energy rules meet resistance from two jurisdictions at their first test

Prime Minister Anthony Albanese's proposed environmental and energy controls on a A$150 billion ($105 billion) data center pipeline went before federal, state and territory energy ministers on Tuesday, their first hurdle since the framework was unveiled (AIPD's July 15th edition), per Bloomberg. Queensland and the Northern Territory declined to back most of the proposed measures, including a national set of regulations. The rules would require operators to match their electricity use with investment in new renewable generation. Dylan McConnell, an energy researcher at the University of New South Wales, said data centers can be built far faster than wind or solar projects. He said gas turbine orders now carry waits of at least three years, largely because of demand from U.S. data centers. Load from data centers is forecast to rise from about 2% of consumption on Australia's main grid to roughly 10% within 10 years, per ABC News.

Read at Bloomberg ↗ • Read at ABC News ↗

Indian state police block more than 400 accounts in an AI disinformation campaign traced partly to Pakistan

Maharashtra Cyber, the Indian state's cybercrime agency, identified 429 social media handles and posts on X, Instagram, Facebook and YouTube used to spread disinformation during recent protests and has had more than 400 of them blocked or removed, per Moneycontrol. Additional Director General of Police Yashasvi Yadav said at least 100 of the accounts originated overseas, primarily in Pakistan along with some West Asian countries and Malaysia. Investigators said AI tools were used to clone the voice of former Union education minister Dharmendra Pradhan, digitally exaggerate crowd sizes and falsely attribute statements to public figures. Police said more than 140 of the identified posts and accounts involved AI generated or AI assisted content, and that no first information report, the filing that formally opens an Indian criminal case, has been registered so far.

Read at Moneycontrol ↗