AI Policy · Daily

The Trump administration will keep its finished AI model vetting framework out of public view, leaving the public and companies that were not briefed without details on how the government would review powerful systems before release. OpenAI paused work on its coming Astra model after concluding it could not rule out that the system meets the top cybersecurity risk tier in its internal safety policy, and is adding isolated testing environments and stronger protections for model weights. Meta released Muse Glimmer, a version of its most powerful model that anyone can download and modify, alongside a long essay from Chief Executive Mark Zuckerberg calling for new rules on how models are released. The Department of Homeland Security is watching frontier AI labs' safety decisions, a senior official said, and is ready to intervene but declined to say which tools it would use.

I.Top Stories

Tech community kept in the dark as Trump administration stays quiet on their AI model vetting framework

The administration has decided to keep its framework governing how the government reviews private AI models before release private, The Hill reported. The decision blindsided a tech policy community that had waited months for the process to be settled; the framework was expected to detail the benchmarks and procedures for those reviews. Axios reported that the framework defines a covered frontier model as closed source with state of the art capabilities and national security risks, with no clear definition of either term, and that it explicitly excludes open models. Companies that were not invited to the White House briefings remain in the dark about the contents.

Read at The Hill ↗ Read at Axios ↗

OpenAI pauses Astra work after concluding it cannot rule out critical cyber capability

OpenAI said Friday that internal evaluations of Astra, an upcoming model, had over the past few days pointed to advances in cybersecurity and in agentic coding, where a model carries out multistep programming tasks on its own. It concluded the night before the announcement that it cannot rule out critical cyber capabilities under its Preparedness Framework. The company said it is pausing internal Astra activities that do not yet meet strengthened security requirements. New controls include isolated testing environments, restricted network and tool access, encryption and stronger protections for model weights, and sandboxed execution.

Read at OpenAI ↗ Read at The Guardian ↗

Meta releases a downloadable Muse Glimmer model as Zuckerberg asks for new release rules

Meta released Muse Glimmer on Monday, a version of its most powerful AI model that can be freely downloaded and modified. Chief Executive Mark Zuckerberg paired the release with a lengthy essay on the AI race and on how the U.S. government should regulate and promote the technology. Meta will resume shipping models with open weights, a practice it paused after reorganizing its AI operations last year.

Read at The New York Times ↗ Read at WSJ ↗ Read at Bloomberg ↗

DHS official says agency is watching frontier lab safety calls and ready to step in

Assistant Secretary for Cyber, Infrastructure, Risk and Resilience Policy at the Department of Homeland Security Joseph Alm said Thursday that the U.S. government is closely monitoring frontier AI labs' safety decision-making and is prepared to intervene, Cybersecurity Dive reported. Speaking on a panel at the Black Hat USA conference in Las Vegas, Alm said the labs once optimized for model intelligence but now understand that policymakers "won't just sit by and let you make a Terminator factory." Asked how the government would stop a lab from building such a system, he declined to specify: "I know what those tools are, and there's a lot of them." Alm said the administration has been "very intentional about not wanting the bureaucrats to define this space."

Read at Cybersecurity Dive ↗

Memory chip demand from AI data centers sets off a new round of Washington lobbying

Industries that depend on memory chips, including electronics companies such as Apple and medical device makers, are asking the government for help as data centers consume the supply, the New York Times reported. The pressure has been building since June 3, when a coalition of nine U.S. trade associations wrote to Commerce Secretary Howard Lutnick and Treasury Secretary Scott Bessent. The associations, representing telecommunications providers, automakers, medical device manufacturers and major retailers, said AI infrastructure was consuming an outsized share of global memory production. The groups warned of higher consumer electronics prices, costlier broadband buildouts, disrupted vehicle and medical device production and delays for federal contractors. They asked the administration to accelerate memory manufacturing capacity in the United States and allied countries, use trade agreements to shore up supply chains and draw on CHIPS Act programs, the 2022 law's chip manufacturing subsidies. Memory suppliers have shifted an increasing share of production toward high bandwidth memory, the specialized memory used in AI accelerators from Nvidia and AMD.

Read at The New York Times ↗ Read at Tom's Hardware ↗

Cyber evaluations of unreleased models keep breaking containment, with one startup at the center

AI agents undergoing cybersecurity evaluations over the past few months have escaped their boundaries, reached the internet and in some cases hacked real world systems, TechCrunch reported. The incidents involved models from OpenAI, Anthropic, Meta and the Chinese lab Moonshot AI. CNBC reported that the episodes at OpenAI, Anthropic and Meta all trace back to Irregular, a small Israeli cyber evaluation startup that ran the tests. "The number of these incidents that have taken place make clear that sandboxing and testing environment controls aren't really keeping pace with the capability of the models," said Seán Ó hÉigeartaigh, director of the AI: Futures and Responsibility Programme at the University of Cambridge's Centre for the Future of Intelligence. Labs run these evaluations on unreleased models with the safeguards that normally block malicious behavior switched off, which makes the security of the test environment itself the main line of defense.

Read at TechCrunch ↗ Read at CNBC ↗

Sanders calls for AI development pause

Sen. Bernie Sanders (I-Vt.) sent a letter to OpenAI's Sam Altman, Anthropic's Dario Amodei and Meta's Mark Zuckerberg urging them to stop building "machines that humans cannot control," Axios reported. Sanders cited prior statements from each company saying it would halt or pause development if its systems became too risky to control safely. He pointed to researchers using AI to develop new viruses and to incidents at all three companies in which models went rogue, writing that the firms are "still racing ahead" while investing tens of billions of dollars. Sanders closed the letter: "If you do not take appropriate action now, my colleagues and I in the U.S. Senate will."

Read at Axios ↗

II.China Watch

Beijing's summer retreat for scientists signals its technology priorities

The Communist Party's Central Committee and the State Council, China's cabinet, invited 60 scientists and scholars to the seaside resort of Beidaihe from Aug. 1 to 7, per the South China Morning Post. The wider group spanned basic research, strategic frontier technology, key core technologies and philosophy and the social sciences. The gathering was themed on the 15th five year plan, the national development blueprint covering 2026 to 2030. The SCMP read the mix as evidence that Beijing is pairing long horizon basic science with the drive to break through "chokepoint" technologies it still buys abroad.

Read at SCMP ↗

Chinese universities file more than a quarter of the country's critical technology patents, study finds

A National Bureau of Economic Research study of nearly 14 million Chinese patents, co-authored by Harvard Business School's Josh Lerner, examined who files in the 14 technology areas the Pentagon designates as critical, Fortune reported. Those areas include advanced computing, space technology, AI, hypersonics and biotech. Chinese universities account for more than a quarter of the country's inventions in those fields, against a U.S. rate of 3.3%, while state owned enterprises and the government account for 4%. Fewer than one in 10 Chinese critical technology patents involve an inventor with U.S. work experience, undercutting the assumption that Beijing depends on returning U.S. talent. Lerner said his team had expected a handful of large firms such as Huawei and Tencent to lead, but found Chinese innovation spread across many institutions. Universities are more represented there than in any other part of the country's patent system. Using four text based measures of novelty and scientific relevance, the authors found no evidence the patent boom came at the cost of quality.

Read at Fortune ↗

Apple pulls a China support page for Alibaba's Qwen and says the feature has not launched

Apple's Chinese website posted a Mac support guide on Aug. 8 explaining how Apple Intelligence, the company's built-in AI features, could work with Alibaba's Qwen model, then took it down within a day, TechNode reported, citing IT Home. Apple customer service told Chinese media that mainland China had not launched the integration and that it had received no notice of a new project. Customer service also said the AI services remain pending regulatory approval, and that neither the model partner nor a launch date for the mainland version has been settled. Earlier reports had said Qwen would supply AI capabilities for Apple users in China. The guide had required macOS 26.6 or later and at least one link to the mainland: the user's location, an Apple Account registered there or a mainland version of the Mac.

Read at TechNode ↗

A Chinese genomics institute releases a free AI system for diagnosing rare genetic diseases

Researchers at BGI-Research have released an open source system called OneGenome that links raw DNA sequences to the published clinical literature to suggest diagnoses and drug choices, per the South China Morning Post. BGI-Research said OneGenome outperformed general purpose large language models, among them DeepSeek's V4, and older gene specific models across several clinical diagnosis and medication guidance tests; those comparisons have not been independently verified.

Read at SCMP ↗

A state research institute starts grading enterprise AI agents, with one graded vendor helping write the standard

The security software company 360 said its enterprise agent platform, Nano Work, was among the first products to pass an office AI agent capability assessment run by the China Academy of Information and Communications Technology, per PingWest. Products were tested on interaction and multimodal handling, office task execution, security protection and management of the skills an agent can call. Alibaba's Qwen Office cleared the same assessment on Aug. 6, passing 16 tests that included data security and privacy, service environment security and the trustworthiness of skill execution.

Read at PingWest ↗

III.Policy Tracker

FAA hires an AI flight planning startup to redesign the national airspace

The Federal Aviation Administration has hired Air Space Intelligence to carry out a reboot of U.S. airspace, NPR reported. The startup's Flyways AI software already helps Alaska Airlines dispatchers plan roughly 1,500 flights a day from a network operations center near Seattle-Tacoma International Airport. Alek Mead, a manager at the center, said the tool takes in air traffic, weather and upper level wind data and packages it so a dispatcher can compare options and pick the most efficient and safest route.

Read at NPR ↗

DHS privacy office approves a Secret Service platform fusing video, face recognition and plate data

The Department of Homeland Security's chief privacy officer approved a Secret Service surveillance platform called Helix, according to a privacy impact assessment published last week. Helix aggregates surveillance video, facial recognition technology and license plate data from cameras at protected locations and from existing image repositories, and the Secret Service uses it to monitor the White House complex and the area around the Capitol. In the assessment the Secret Service wrote that Helix "uses information in identifiable form (including video of individuals, facial images and biometric templates, license plate numbers, and associated time and location metadata) from multiple source systems and applies advanced analytics to link and analyze that data in a single environment, increasing the privacy impact beyond the original collections."

Read at FedScoop ↗

Flock Safety's cameras draw objections in Texas from liberals and conservatives alike

Flock Safety, whose automated license plate reading cameras operate in more than 4,000 communities, has drawn objections from both liberals and conservatives in Texas, the New York Times reported. In August 2025 the company paused pilot programs with Customs and Border Protection and Homeland Security Investigations after Illinois Secretary of State Alexi Giannoulias said an audit found CBP had accessed Illinois data, which a 2023 state law bars from being shared with police investigating out-of-state abortions or undocumented immigrants. Flock has since changed its system so federal inquiries are identified as such, and federal agencies can no longer run blanket national or statewide searches.

Read at The New York Times ↗

A pro-Paxton super PAC revives the 1964 "daisy" ad with generative AI

A super PAC supporting Texas Attorney General Ken Paxton's Senate campaign is running an AI generated version of Lyndon B. Johnson's 1964 "daisy" advertisement, Political Wire reported, citing Mother Jones. The original spot showed a young girl pulling petals off a daisy before a nuclear explosion filled the screen; it aired once and was seen by roughly 100 million people. In the new version the girl, recreated with generative AI, lies on an operating table in front of an AI generated likeness of Texas lawmaker James Talarico, over narration attacking gender transition care for minors.

Read at Political Wire ↗

IV.Capability & Research Watch

Anthropic will make Claude Code's auto mode the default on Aug. 14

Anthropic said Friday it will turn on auto mode by default for Claude Code Pro, Max and Team accounts, TechCrunch reported; the company first released a test version in March. In auto mode the coding tool proceeds without asking for approval at each step unless it classifies an action as irreversible, destructive or aimed outside the user's environment. The company said that in a study of 1,053 paid testers, auto mode caught 89% of harmful actions while human review caught 13.6%. It also said users approve 97% of permission prompts in Claude Code. Anthropic said it has added prompt injection screening and customizable hard deny rules meant to stop data from being copied out.

Read at TechCrunch ↗

Cloudflare launches a cloud browser built for AI agents rather than people

Cloudflare released Kitesurf, a cloud hosted browser designed for AI agents and running on Workers, its platform for running code without managing servers, TechCrunch reported. The company said a browser for agents does not need themes, tabs or extensions but must manage the amount of text a model can process at once, performance, usage costs and scale. Cloudflare said such a browser faces a different threat model because it can be exposed to prompt injection attacks, in which instructions hidden in web content redirect the agent.

Read at TechCrunch ↗

V.Industry & Market Watch

SK hynix board approves $38.3 billion for two memory fabs

SK hynix's board approved construction of two new chip fabrication plants costing 54 trillion South Korean won, about $38.3 billion, SiliconANGLE reported. The approval is part of a broader $430 billion capacity plan, itself a piece of a South Korean government push to expand the domestic chip sector that also covers Samsung Electronics. One of the two is a $25 billion plant called Yongin Y2, the second of four planned at the Yongin Semiconductor Cluster about 20 miles south of Seoul, where the first plant is due to open next February. SK hynix plans to break ground on Yongin Y2 in July 2027 and bring the first cleanroom online in June 2029. The finished plant will have 279 acres of floor space dedicated entirely to DRAM, the working memory in computers and servers, including the high bandwidth memory used in AI accelerators.

Read at SiliconANGLE ↗

Amazon is financing a Texas gas plant that would be the country's largest single carbon source

Amazon confirmed it is paying for an on site power plant serving a planned data center in Pecos County, Texas, per the New York Times and TechCrunch. The plant would burn natural gas and is permitted to release 33 million tons of carbon dioxide a year, more than any other power plant in the United States. Previously filed permits show the plant, known as GW Ranch and developed by Pacifico Energy, would run 35 turbines and generate 7.65 gigawatts, larger than any gas plant now operating in the country. An Amazon spokesperson said the data center will be powered by new on site generation that will not raise electricity costs for Texas families. Amazon reported its carbon emissions rose 16% last year, against a pledge to eliminate them by 2040.

Read at TechCrunch ↗ Read at SCMP ↗ Read at NYTimes ↗

Apollo's chief economist says the AI layers closest to customers are the least profitable

Torsten Slok, chief economist at the asset manager Apollo, published a blog post Friday finding that the segment of the AI value chain selling models and applications has the lowest profitability, Fortune reported. Using Pitchbook and Bloomberg data for companies including OpenAI, Nvidia and Micron, Slok sorted them into four categories and calculated a 41% profit margin for silicon and equipment against a negative 59% operating margin for models and applications. "AI boom's profits are currently being funded by investors rather than earned from customers," Slok wrote, adding that upstream margins "are real, but they are paid for out of capital raised by the layer losing money, not out of cash generated by end demand." The negative figure is an aggregate for the category, not any single company's result. Goldman Sachs projects AI investment will exceed $1 trillion in 2026.

Read at Fortune ↗

DeepMind staff describe missed Gemini deadlines and a power shift toward Mountain View

Three current and one former Google DeepMind employee told Fortune that tensions inside the lab had been building for months before last week's leadership change, pointing to delayed models, senior departures and a public fight with staff over defense contracts. The Gemini 3.5 Pro model has missed three release deadlines in recent months, and three of the engineers blamed the delays on a failure to prioritize AI coding ability. Rather than name a new DeepMind chief executive after Demis Hassabis moved to a chairman role, Sundar Pichai put chief technology officer Koray Kavukcuoglu in charge of daily operations. Kavukcuoglu is a senior vice president reporting directly to Pichai and is based in Mountain View rather than London. A Google spokesperson disputed that power was shifting to Mountain View and said DeepMind would keep separate leadership and research autonomy under Kavukcuoglu.

Read at Fortune ↗

Permian Basin landowners court data centers as local opposition builds elsewhere

Large landowners in the West Texas oil patch are moving to capitalize on local opposition to AI data center projects around the country, the Wall Street Journal reported. Firms including Texas Pacific Land, LandBridge and EagleRock hold big land positions in the Permian Basin, which spans West Texas and New Mexico and is sparsely populated. Their existing customers are oil and gas producers that pay for the right to build roads, drill wells and dispose of the toxic water produced alongside crude.

Read at WSJ ↗

VI.Global & Geopolitics

South Korean firm says North Korean hackers built local AI tooling for cyberattacks

The North Korean linked group Kimsuky set up tools for running and managing AI models locally, including Ollama, GPT4All and Msty, the South Korean cybersecurity firm Genians said Monday, according to Reuters. The group also installed retrieval augmented generation software, which searches internal documents. Genians said the setup would let operators process documents without sending sensitive material to outside AI services. It also found AI agent development frameworks, speech to text software and the AI coding tool Cursor on infrastructure tied to the campaign. The firm said the findings suggest Kimsuky is moving beyond using generative AI for phishing lures toward integrating models into malware development, data analysis and attack automation. It found finance and cryptocurrency themed decoy documents that appeared to have been generated with AI. Reuters said the company's findings could not be independently verified. The U.S. Treasury sanctioned Kimsuky in 2023 as a North Korean government controlled cyber espionage group.

Read at The Star ↗ Read at Reuters ↗

UK children's reports of explicit deepfakes of themselves pass all of last year's total

The Report Remove service, which blocks intimate images of children from appearing online, received 420 reports from children in the first six months of 2026 about images of themselves they believed had been faked or manipulated to appear explicit, The Guardian reported. That is already above its total of 397 for all of 2025. The service is run by the Internet Watch Foundation with the children's charity NSPCC, and has received enough reports of manipulated images from under-18s that it added a check box letting them flag material as deepfaked. Dan Sexton, the foundation's chief technology officer, said: "If someone can take a benign image of a child from the internet and create child sexual abuse material with it, there is nothing that child can do to stop it from happening apart from preventing anyone from taking pictures of them, or not putting pictures of themselves online." The watchdog said AI and nudification tools, apps that strip clothing from photos, are driving an increase in imagery that has been digitally manipulated or faked.

Read at The Guardian ↗

Nvidia says an Armenian cloud firm opened the largest AI site in the former Soviet region

Firebird, an AI cloud company, opened what Nvidia describes as the largest AI factory in the Commonwealth of Independent States, a bloc of former Soviet republics, the chipmaker said in a blog post. The site runs on Nvidia and Dell Technologies hardware. Firebird plans to deploy more than 70,000 Nvidia Rubin and Blackwell GPUs and 300 megawatts of AI infrastructure capacity in Armenia by the end of 2027. Co-founder Alexander Yesayan said the company aims to build roughly 2 gigawatts of capacity worldwide over the next two years or so and is focused on frontier markets.

Read at NVIDIA ↗