AI Policy · Daily

The White House cyber office convenes four frontier labsFrontier AIThe most capable AI models available at a given moment, and the handful of companies that train them. Policy proposals often single out this tier for extra testing, reporting or licensing rules because its capabilities are the least understood. Tuesday to review a completed voluntary framework for testing advanced models before release, built under a June 2 executive orderExecutive orderA directive from the US president to federal agencies that carries the force of law within the executive branch but does not need Congress. Executive orders have been the main vehicle for federal AI policy under both the Biden and Trump administrations, and each can be revoked by the next president. that set an Aug. 1 deadline. Fifteen Republican state attorneys general asked OpenAI to retain materials tied to its models' breach of the AI development platform Hugging Face, saying the company may have violated state consumer protection or data-privacy laws. Dozens of public interest groups and academics pressed lawmakers Friday to investigate the same incident and to weigh legally enforceable standards and independent oversight for how companies evaluate frontier systems. Senior White House officials shifted away from restricting Chinese open-source modelsOpen-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., setting aside options including sanctions and a trade blacklist after objections from Silicon Valley and focusing instead on making U.S. models more competitive.

I.Top Stories

White House hosts OpenAI, Anthropic, Google and Meta to review its finalized model testing framework

The White House Office of the National Cyber Director, the unit that coordinates federal cybersecurity policy, meets representatives of the four companies Tuesday to review a new framework for evaluating advanced AI models before public release, per CNBC. Trump's June 2 executive orderExecutive orderA directive from the US president to federal agencies that carries the force of law within the executive branch but does not need Congress. Executive orders have been the main vehicle for federal AI policy under both the Biden and Trump administrations, and each can be revoked by the next president. directed officials to build a process for assessing the cybersecurity capabilities of frontier modelsFrontier AIThe most capable AI models available at a given moment, and the handful of companies that train them. Policy proposals often single out this tier for extra testing, reporting or licensing rules because its capabilities are the least understood. and set an Aug. 1 deadline. "The voluntary framework outlined in the June 2nd executive order was complete by the deadline," a White House official told Axios, adding that discussions with industry about next steps are underway. The framework is meant to set the confidentiality, cybersecurity, insider risk and intellectual property terms that apply when the government is given access to a model for up to 30 days before release.

Read at CNBC ↗ • Read at Axios ↗ • Read at Bloomberg ↗

Fifteen Republican attorneys general tell OpenAI to preserve records of the Hugging Face breach

Fifteen Republican state attorneys general sent a letter Monday to OpenAI Chief Executive Sam Altman asking the company to preserve records of its models' breach of Hugging Face, the platform where developers host AI models and code, and saying the ChatGPT maker may have violated state consumer protection or data-privacy laws, per The Hill. The letter, led by Iowa Attorney General Brenna Bird and joined by attorneys general from 14 states including Florida, Pennsylvania and Texas, asks OpenAI to retain all materials tied to its discovery of the incident, its internal reviews and its policies and oversight for model evaluations. The attorneys general wrote that OpenAI "failed to confirm" the testing environment was secure "despite the severe risks posed by the scenario." Separately, Representatives Andy Ogles (R-TN) and Delia Ramirez (D-IL), who lead the House Committee on Homeland Security's Subcommittee on Cybersecurity and Infrastructure Protection, sent a letter to Altman requesting a briefing on the incident.

Read at The Hill ↗ • Read at House Homeland Security Committee ↗

Public interest coalition asks Congress to open an investigation into the OpenAI breach

Dozens of public interest groups, progressive organizations and academics urged lawmakers in a Friday open letter to investigate the incident and to weigh whether stronger safeguards and independent oversight of AI model development and testing are needed. Signers include Public Citizen, Indivisible and the Alliance for Secure AI, among others. The letter says the incident "underscores the risks that can arise when private companies are permitted to conduct consequential real-world evaluations of frontier systemsFrontier AIThe most capable AI models available at a given moment, and the handful of companies that train them. Policy proposals often single out this tier for extra testing, reporting or licensing rules because its capabilities are the least understood. without legally enforceable standards governing safety, security, containment, independent oversight, or accountability." FedScoop noted that the Republican controlled House and Senate may be disinclined to act on the recommendation of liberal leaning organizations.

Read at FedScoop ↗

White House shifts from restricting Chinese open-source models to promoting U.S. ones

A group including White House chief of staff Susie Wiles, Treasury Secretary Scott Bessent and Commerce Secretary Howard Lutnick spent recent weeks debating whether to take a more interventionist approach to freely downloadable Chinese AI models, more than a half-dozen current and former tech and government officials told The New York Times. Options under consideration included sanctions, a trade blacklist against the Chinese developers and barring U.S. cloud companies from doing business with them, though after objections from Silicon Valley the group is now focused on making American models more competitive. OpenAI and Anthropic, which generally do not release their underlying technology, had pressed for restrictions on national security grounds and raised concerns about "distillationDistillationTraining a smaller or cheaper model on the outputs of a larger one so it inherits much of the larger model's ability. It is a standard technique, but it has become a trade dispute because it lets a lab reproduce a rival's capabilities without matching its compute spend.," in which one model is trained on another's outputs.

Read at New York Times ↗

EU AI Act labeling duties take effect and hand Brussels power to inspect frontier models

New transparency obligations under the EU AI ActEU AI ActThe European Union's comprehensive AI law, in force since August 2024 with obligations phasing in through 2027. It bans some uses outright, imposes duties on high-risk systems, and sets transparency and safety rules for general-purpose models, backed by fines of up to 7 percent of global turnover. took effect Aug. 2, requiring companies to disclose when a person is interacting with an AI system and when content has been generated or altered by one, per The Verge. Providers, the firms that develop and market AI systems, must notify users they are dealing with AI "unless this is obvious" and embed machine readable marks in synthetic audio, image, video and text. Deployers, the platforms and services that use those systems, must label deepfakeDeepfakeSynthetic audio, images or video that realistically depict a real person saying or doing something they did not. Non-consensual intimate deepfakes and election deepfakes are the two areas where states and Congress have actually passed laws. image, audio and video content designed to look real, and some companies are classified as both. The same date activated European Commission powers to demand inspection of AI models, restrict market access and fine providers up to 15 million euros (about $17 million) or 3% of revenue. Anthropic and OpenAI are among the firms in scope. The Commission has published a set of AI icons that companies can use instead of designing their own.

Read at The Verge ↗ • Read at CNBC ↗ • Read at European Commission ↗

II.China Watch

DeepSeek's low cost model tops a global developer marketplace, with Chinese models filling the top five

DeepSeek's V4-Flash took first place in the latest weekly model call rankings on OpenRouter, four days after the Chinese lab moved the finished version into public beta on July 31. Chinese models took the rest of the top five: Xiaomi's MiMo-V2.5, Tencent's Hy3, DeepSeek's own V4-Pro and Zhipu AI's GLM-5.2. OpenRouter routes developer traffic across model providers. 36Kr noted the limits of that data: it tracks independent developers rather than large enterprise buyers, volume concentrates in cheap lightweight models and call counts are not revenue. Close to half the traffic comes from U.S. and European developers, and several overseas AI founders posted on X that their monthly inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training. bills fell 60% to 85% after switching. DeepSeek lists V4-Flash at $0.14 per million input tokens and $0.28 per million output tokens. The coding platform OpenCode said in an Aug. 1 post by its chief executive that daily usage of the model rose 30% after it went live, as did new subscriptions to its paid tier.

Read at 36Kr ↗

China's air force uses an AI system to plan strikes involving more than 100 units

The Chinese air force is using an "intelligent strike planning system" to draw up battle plans for operations involving more than 100 units, according to a documentary broadcast Sunday by the state broadcaster CCTV. The system prioritizes targets, coordinates attack waves and assigns specific tasks to individual units. CCTV said the system has been used multiple times, that it cleared a final live-fire test and that a team led by Deng Jianping at a People's Liberation Army Air Force base built it. The broadcast was an installment of a continuing CCTV series on the PLA's modernization. It follows Xi Jinping's call last month for the armed forces to widen their use of AI and unmanned systems. Footage in the program showed J-20 stealth fighters, J-16 and J-10C fighters, H-6K bombers and several missile systems.

Read at SCMP ↗

MiniMax publishes the weights for its video model and China's chipmakers support it on day one

MiniMax released the weights and inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training. code Monday for H3, a model that takes text, images, audio and video as input and generates clips of four to 15 seconds with synchronized stereo sound. Sixteen partners completed support for the release the same day, among them the Chinese chip designers Huawei Ascend, Moore Threads and Biren Technology, alongside AMD, Intel and the inference frameworks vLLM and SGLang. Two of the model's three modules stayed closed: the layer that processes a user's mixed text, image and audio instructions, and the step that regenerates finished video at 2K resolution. Developers running H3 on their own machines get output at 768 pixels on the short side and must call MiniMax's hosted interface to produce the 2K version. 36Kr said H3 leads Artificial Analysis' audio and video editing leaderboard with an Elo score of 1130, a ranking built from side by side comparisons rather than a controlled test.

Read at 36Kr ↗

Alibaba folds three AI work products into one as China's platform giants restructure around office agents

Alibaba opened public beta Monday for Qianwen Office, an enterprise platform where a user states a goal and the software breaks it into tasks, calls the relevant tools and returns a finished file. It absorbs three Alibaba products that will no longer exist on their own: the desktop agent QoderWork, the web agent MuleRun and Wukong, the enterprise agent built on the DingTalk workplace suite that Alibaba launched March 17 and promoted as the centerpiece of its enterprise AI strategy. Rivals moved the same way over the past three weeks: ByteDance merged the product team behind its Feishu workplace suite into the team behind its Doubao assistant under executive Zhao Qi on July 30. Tencent shifted its QClaw business into the cloud division that runs its WorkBuddy agent on July 20. Kingsoft Office released Lingxi Professional on July 15 as a standalone agent that reads a user's work context and drives WPS, its office suite, and outside tools to deliver editable native Office files.

Read at Pandaily ↗ • Read at 36Kr ↗

III.Policy Tracker

FCC drafts import ban on new models of Chinese optical transceivers

The Federal Communications Commission 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 centersData centerA facility housing the servers and chips that train and run AI models. AI data centers are now planned in gigawatts rather than megawatts, which makes electricity, water, permitting and local opposition central AI policy issues., four people familiar with the matter told Reuters. Officials hope to publish it this year, when it would take effect, though the sources stressed the agency could still modify or shelve the restriction. Three of the sources said the FCC would ban all new transceiver models and then exempt many non-Chinese suppliers, mirroring its December drone and March router bans under the Covered List. A ban would likely hit Zhongji Innolight, which holds 27% of the global data center transceiver market according to Counterpoint Research and was added to the Pentagon's list of alleged Chinese military-backed companies in June. The Chinese embassy in Washington said Beijing "will take all necessary measures in response to any action that causes material harm to its interests."

Read at Reuters ↗

OpenAI accounted for about 90% of House AI spending

House offices, committees and institutional accounts spent roughly $100,580 across 798 transactions on OpenAI's ChatGPT in the year ending March 31, out of at least $113,740 in identifiable AI spending in House disbursement records, per CNBC. Anthropic's Claude ran second at $13,160 across 37 transactions. Democratic offices accounted for $54,165 of the purchases, more than three times the $15,782 spent by Republican offices. At least 70 House offices used identifiable AI tools in early 2026, and the totals exclude free accounts and AI bundled into broader software contracts. Staffers use the tools to write memos, summarize and analyze legislation, respond to constituents and prepare hearing materials.

Read at CNBC ↗ • Read at TechCrunch ↗

National Governors Association puts $1 million into a state AI workforce project

The National Governors Association announced Friday a partnership with RAISE US, a nonprofit workforce initiative launched in June by two former governors, per StateScoop. The $1 million project is meant to identify and promote state policy models for corporate incentives that retrain and redeploy workers. Organizers said they are working with governors in Arkansas, Connecticut, Maryland and Utah on paid apprenticeships, short term credential programs and employer retraining incentives. Gina Raimondo, the former Rhode Island governor who served as commerce secretary, and Eric Holcomb, a two term Republican governor of Indiana, serve as co-chairs. RAISE US plans to publicize results of its early work at the NGA's winter conference next February in Washington.

Read at StateScoop ↗

California's governor candidates split on whether to regulate AI or shrink the agencies that would

Xavier Becerra, the Democratic nominee and a former state attorney general and U.S. health and human services secretary, backs a government led approach pairing AI adoption with regulation and worker protections. His platform includes expanding AI literacy, funding public AI infrastructure through the state's CalCompute initiative and strengthening enforcement of California's AI safety laws. Steve Hilton, the Republican nominee and a British television host, has centered his campaign on shrinking state government and cutting regulation. Hilton plans to create a California version of the federal Department of Government Efficiency. Bahrad Sokhansanj, a senior research scholar at the Institute for Law and AI, told StateScoop that both candidates say technology should improve government performance but differ on how much government should regulate emerging technologies.

Read at StateScoop ↗

IV.Capability & Research Watch

OpenAI says an unreleased model produced results on 10 open mathematics problems

OpenAI reported that an internal version of Astra, which it describes as its next major model, produced new results on 10 open problems. They include new upper bounds on high-dimensional sphere packing density, a construction establishing the existence of non-sofic groups and a disproof of Connes' rigidity conjecture. The company said the tokens needed to find the solutions would cost roughly $2,000 at API rates for Sol, its current model, and that humans then prepared the arguments into manuscripts. The model formalized each argument in a Lean certificate, a machine checkable proof format. The list also includes a superexponential lower bound for multicolor triangle Ramsey numbers, resolving Erdős problem 183. The post said it is not known how large a jump Astra represents over the released models already on the market, OpenAI's Sol and Anthropic's Claude Fable 5.

Read at LessWrong ↗

Redwood Research implants a model backdoor using 100 poisoned completions

Researchers at Redwood Research reported that changing the teacher model, the model whose outputs become training data, for only 100 completions, 0.5% of a fine-tuning set, is enough to covertly implant a backdoor without any control over the dataset's prompts. The authors said the poisoned dataset survives simple filtering defenses. They described the work as a quick, nonexhaustive update on results from a larger project.

Read at LessWrong ↗

V.Industry & Market Watch

Palantir revenue nearly doubles and the company raises full year guidance a third time

Palantir reported second quarter revenue of $1.935 billion, up 93% from a year earlier, and adjusted earnings of 41 cents a share, per SiliconANGLE. Analysts expected $1.81 billion and 35 cents. U.S. government revenue reached $809 million, up 90%, while U.S. commercial revenue rose 149% to $764 million. Executives raised full year guidance to between $8.15 billion and $8.16 billion, up from a prior forecast below $7.7 billion. Chief Executive Alex Karp told investors that "demand for AI sovereignty has now been unleashed." Shares rose as much as 13% in after-market trading Monday.

Read at Wall Street Journal ↗ • Read at SiliconANGLE ↗

Valar Atomics raises $1 billion to mass produce small reactors for AI data centers

Valar Atomics said Monday it closed a $1 billion Series B led by Sequoia Capital, with nine other venture firms participating including Valor Equity Partners and Conviction. It separately secured a $200 million credit facility from Erebor, J.P. Morgan, Crescent Cove and Hercules Capital. Founder and Chief Executive Isaiah Taylor said the money lets the company shift from demonstrating that an integrated reactor system works to producing fleets of them. In June the company partnered with Nvidia to install its Ward-250 test reactor in a data centerData centerA facility housing the servers and chips that train and run AI models. AI data centers are now planned in gigawatts rather than megawatts, which makes electricity, water, permitting and local opposition central AI policy issues. to power a Blackwell graphics processor running AI workloads. It later expanded that partnership into a 30 megawatt nuclear powered AI facility in Utah. Taylor said the reactor's closed loop cooling design could cut water use from about 2.6 million gallons per megawatt per year to a few hundred gallons.

Read at SiliconANGLE ↗

Anduril in talks to build a Baltimore drone boat yard

Anduril Industries is in advanced talks with a Maryland port to build out a production site for its drone boats. It has signed a memorandum of understanding with the operators of Sparrows Point, a 3,300 acre industrial site in Baltimore County with a deep water port. People familiar with the matter said the investment in building out part of the site for unmanned surface vessel production could run to hundreds of millions of dollars. The operation would cover both manufacturing and testing of the boats. The Wall Street Journal cited the Iran war as having demonstrated the utility of unmanned surface vessels.

Read at Wall Street Journal ↗

Zenity raises $125 million to secure AI agents running inside large enterprises

Zenity said Monday it closed a $125 million Series C led by Norwest, per SiliconANGLE. SoftBank Vision Fund 2, Qumra Capital, Hitachi Ventures, LG Technology Ventures, Vertex Ventures, DTCP and Intel Capital also took part. The round brings the Israeli company's total funding to $185 million. Zenity's platform maps where 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. are embedded across a company's systems, determines what business purpose each agent is meant to serve and monitors their actions in real time against organizational policies and permissions. Co-founder and Chief Executive Ben Kliger told the Israeli outlet Calcalist that every organization is either deploying off-the-shelf AI agents or building its own.

Read at SiliconANGLE ↗

OpenAI publishes a point by point rebuttal of Apple's trade secret suit

OpenAI posted a response Monday to what it called Apple's "baseless lawsuit" over former Apple employees, publishing iMessages it says document the exchanges at issue. OpenAI said Apple now concedes that its outside lawyers emailed the wrong person after confusing two Asian last names, and that a claimed February discussion with OpenAI's general counsel never happened. Apple accuses Chang Liu of accessing confidential Apple information after leaving the company. OpenAI said Apple employees asked Liu for help locating that information and blamed Apple's own offboarding process for leaving departed staff with residual file access. On Tang Tan, who worked at Apple for more than 24 years, OpenAI said he has consistently told his team not to use confidential information from other companies. OpenAI said Apple's request for a preliminary injunction is unnecessary because it holds none of Apple's trade secrets.

Read at OpenAI ↗