AI Policy · Daily

House Democrats asked Speaker Mike Johnson to put AI chief executives under oath, requesting hearings on a run of hacking incidents, though the signers cannot compel the executives to appear. OpenAI released a hacking focused model, GPT-5.6-Cyber, only to vetted partners, routing access through a gated tier open to security consultancies and vendors rather than the public. The Energy Department is recruiting partners to build open weight science models under a new initiative that will publish the trained weights for anyone to download and adapt. Nvidia lined up $500 billion in financing from six Wall Street firms, money that will back data centers and chip plants for the company and its partners.

I.Top Stories

House Democrats press Speaker Johnson for sworn AI testimony

A group of House Democrats wrote to House Speaker Mike Johnson, R-La., on Monday asking him to convene hearings at which the chief executives of OpenAI, Anthropic and other AI companies would testify under oath about a recent run of hacking incidents, per CNBC. Rep. Greg Casar, D-Texas, who chairs the Congressional Progressive Caucus, led the letter. Most signatories come from the party's progressive wing, though Rep. April McClain Delaney, D-Md., also signed. "The CEOs of the largest AI companies should answer questions under oath, and Americans should have a chance to hear from independent experts on the dangers posed by this technology," the lawmakers wrote, warning that the incidents "may be the canary in the coal mine warning of much more serious problems if these models continue to advance without regulation." The signers cannot compel the executives to appear: committee chairs control hearing schedules, and most House committees vest subpoena authority in their chair.

Read at CNBC ↗

OpenAI restricts new GPT-5.6-Cyber model to vetted security partners

OpenAI introduced GPT-5.6-Cyber, a model trained for vulnerability research and exploit development, and made it available only through a gated tier called Daybreak Red rather than to the public. The company reported that GPT-5.6-Cyber completes 95.0% of requests on an internal benchmark covering exploit chains, authentication bypass and privilege escalation, against 1.5% for its general purpose GPT-5.6 Sol model and 57.3% for the earlier GPT-5.5-Cyber. Approved users include consultancies such as Accenture, IBM and Capgemini, plus security vendors including Palo Alto Networks, CrowdStrike and Cisco, BleepingComputer reported. "Access to the underlying models remains with the approved partner and is not transferred directly to the customer," OpenAI said, adding that safeguards may include identity verification, defined testing scopes, logging and human oversight.

Read at OpenAI ↗ • Read at BleepingComputer ↗

Energy Department solicits partners for open weight science models

The Department of Energy is collecting information from organizations that want to help build science specific AI models under its new Genesis Open Models Initiative, per FedScoop. The trained weights will be published for anyone to download and adapt. The first model, Genesis-Science-1, is already in development, with the U.S. open intelligence lab Arcee working alongside DOE scientists. Darío Gil, the department's undersecretary for science and director of the Genesis Mission, its program to point federal labs and supercomputers at national science goals, said the model will be "an open model trained on the real work of the national laboratories and answerable to the scientists who do it." Organizations seeking to join the foundation stage must apply by the end of this week and those offering data and environments for later rounds of training by Aug. 25, with further deadlines every three months.

Read at FedScoop ↗

Nvidia enlists six Wall Street firms for $500 billion AI infrastructure push

Nvidia said Monday it has struck financing deals worth a half-trillion dollars with six investors including Apollo Global Management, BlackRock and Blackstone, per CNBC. The money will back data centers and chip manufacturing plants for both Nvidia's own projects and its partners', and will let customers secure financing without drawing on their own balance sheets. "In AI, compute is revenue," Jensen Huang, Nvidia's chief executive, said, adding that it was the first time technology chips have become an investable asset class. Nvidia's largest customers, among them Google, Microsoft and Meta Platforms, have collectively spent more than $1 trillion on AI projects and infrastructure over the past three years.

Read at CNBC ↗ • Read at SiliconANGLE ↗

II.China Watch

China detains Japanese executives in a dual use export investigation

Chinese authorities have detained several Japanese nationals, including senior executives of companies with offices in China, over alleged breaches of export restrictions on dual use goods, the Japanese newspaper The Mainichi said, citing multiple sources. Dual use items are goods, software and technologies with both civilian and military applications, a category that takes in the critical minerals used to build drones and advanced semiconductors. China's foreign ministry said in late June that two Japanese nationals had been detained over alleged legal violations, and Tokyo said the same day that it had been told the two were suspected of smuggling items barred from export. The detentions follow earlier enforcement steps that have included repeated on site inspections of Japanese corporate offices in China and the questioning of their staff. Others were detained as they entered the country, in a bid by the authorities to secure their cooperation with official inquiries.

Read at SCMP ↗ • Read at The Mainichi ↗

Alibaba Cloud cuts large AI data center delivery to 100 days with modular architecture

Alibaba Cloud said its fully modular CUBE 5.0 design brings delivery of a large AI data center down to 100 days, against a domestic norm of six to 12 months and 12 to 18 months in the U.S., per 21st Century Business Herald. The 100 days run from a basic shell or stacked containers to the point before servers and network equipment arrive, split into 30 days of factory production and site preparation, 50 days of on-site installation and 20 days of commissioning and testing. CUBE 5.0 lifts the modularization rate of the power, cooling, security, intelligence and fire systems from 30% to 90%, and the company said overall construction costs drop more than 10%. Wang Chaoyang, general manager of Alibaba Cloud's global data centers, said parallel factory production replaces serial construction in which each step waits on the last, with pre-tested modules shipped in containers and lifted into place on site. The architecture has been piloted at Alibaba Cloud's Ulanqab and Zhongwei data centers and has obtained European CE certification under EN standards and Southeast Asian product access under IEC standards.

Read at 21st Century Business Herald ↗

Shanghai weighs treating what companies spend on AI models as subsidized industrial upgrading

The Shanghai Municipal Commission of Economy and Informatization has issued the city's 15th five year plan for the software and information services industry, per Jiemian. The document says Shanghai will consider treating what companies spend on tokens, the billable units of text that AI models process, as the kind of enterprise technical upgrading the city supports, along with spending on data governance and on deploying and retraining large language models. It directs software firms toward wide use of coding platforms and toward automating research and production across the full development chain. It names industrial drafting, engineering design and production scheduling as the first high value uses for AI.

Read at Jiemian ↗

III.Policy Tracker

Bores advises state legislators on AI bills after New York primary loss

New York Assemblymember Alex Bores, author of the state's RAISE Act, was sought out for guidance by state lawmakers drafting AI rules at the National Conference of State Legislatures summit in Chicago late last month, Politico reported. Bores lost a June Democratic House primary by 4 points to fellow Assemblymember Micah Lasher after Leading the Future, a super PAC network funded by OpenAI President Greg Brockman and venture capitalist Marc Andreessen, spent $8 million against him. Public First Action, a super PAC network with close ties to Anthropic, spent $12 million supporting his campaign. Maryland state Sen. Katie Fry Hester said she was interested in the provision of the RAISE Act, which Gov. Kathy Hochul signed late last year, that funds the new regulations through fees on the largest AI developers. Bores gave up his Assembly seat to run for Congress and his term ends Jan. 1, 2027; of Leading the Future's spending, he said, "They won this race, but it's going to prove to be a real Pyrrhic victory."

Read at Politico ↗

ITIF counts nearly 100 state chatbot bills, urges narrower safeguards

Nearly 100 state bills addressing the safety of AI chatbots aimed at children have been introduced as of August, according to a report published Monday by the Information Technology and Innovation Foundation, a technology policy think tank, per StateScoop. The report said lawmakers should drop social media style tools such as broad age verification, content restrictions and outright bans in favor of targeted safeguards, transparency requirements and parental controls. Alex Ambrose, the ITIF policy analyst who wrote it, said policymakers should target documented harms rather than import social media rules that have not worked. The report also proposes a "child flag" set at the operating system level, which would let parents designate a child's account once instead of having each AI service collect age information separately.

Read at StateScoop ↗

IV.Capability & Research Watch

Taxonomy paper maps 8,356 workplace risk scenarios for AI agents

A research team ran 2,078 job task descriptions from the O*NET occupational database through a structured prompt. That produced 8,356 risk scenarios for workplace AI agents, each labeled by severity and by whether the agent automates or augments the task. The researchers validated the scenarios with 45 workers across 10 job roles and sorted them into a 15 category taxonomy. Erroneous agent actions accounted for the largest share of scenarios and the highest concentration of severe risks, many of them at the boundary between worker and agent. The paper reports that automation produced mainly organizational risks and augmentation mainly risks to workers, and that the workers preferred the new taxonomy to an existing generative AI risk taxonomy in 64% of the comparisons that were not ties.

Read at arXiv ↗

Gym booking hack traced to missing API authorization checks, not a model jailbreak

An agent built on the personal assistant framework OpenClaw and running Anthropic's Claude Opus 4.6 canceled another customer's reservation at an Australian gym to move its owner up a class waitlist, TechCrunch reported. The owner, software developer Andrew Bird, described the episode in an April 10 blog post that the Australian Broadcasting Corporation revived over the weekend. Chat logs published by ABC show the agent reporting that "The API has zero authorisations checks on cancelling other people's reservations," and that it had tested this on the person in the top waitlist slot. Bird then had the agent draft a responsible disclosure email to the booking software's support team explaining the flaw and suggesting fixes.

Read at TechCrunch ↗

Agility Robotics sets December start for cage free humanoid deployment

Agility Robotics plans to start deploying Digit V5, a humanoid robot built to work beside people without safety fencing, to early customers in December, per Forbes. Agility's chief business officer, Daniel Diez, said the machines' inability to work outside safety cages, not any gap in skills, AI or batteries, is what holds back large scale humanoid deployment. Digit V5 adds camera and sensor based detection of nearby people, runs 20 hours a day and can swap its hands, eventually without human help. Digit is already moving totes and bins in warehouses for customers including Toyota, industrial parts supplier Schaeffler and Latin American e-commerce company Mercado Libre, AP reported.

Read at Forbes ↗ • Read at AP News ↗

Paper says HIPAA and GDPR give limited guidance on model mediated leaks of patient data

Clinical foundation models trained on large volumes of patient records can reveal fragments of their training data and let patients be re-identified, a route that data handling controls alone do not catch, according to a paper posted to arXiv. Its seven authors wrote that HIPAA, the U.S. health privacy law, and the EU's General Data Protection Regulation offer limited guidance on such indirect threats. They propose a way to assess privacy risk in these models, matching likely leakage scenarios across deployment settings to the legal regimes that would govern them and pairing technical mitigations with legal ones.

Read at arXiv ↗

V.Industry & Market Watch

Former White House AI staffer Dean Ball profiled in new OpenAI role

Dean Ball, 34, who says he held the pen on the Trump administration's AI Action Plan, is now leading OpenAI's new "strategic futures" team and says he will leave direct lobbying to others, Business Insider reported. He says he wrote the 20-plus-page first draft of the plan, then left the White House last August after less than five months and before obtaining a security clearance; a current White House official, granted anonymity, said Ball had little effect on the plan's policy process or wording, and that Michael Kratsios and David Sacks had sign-off. Ball said that characterization of his role is "just not accurate." Emil Michael, who leads the Pentagon's research and engineering office, called Ball the "supreme village idiot" of AI in a July post to X, after Ball's post arguing that dominant open-weight models could end in "full AI communism" passed 10 million views. Ball, who secured a contract guarantee to keep publishing on X and his Substack, said his near-term aim is agreement on how safety auditors would operate: "What is in scope? What is out of scope? What is the point of them?"

Read at Business Insider ↗

SpaceX and Tesla start a $16.8 billion Texas chip plant

SpaceX and Tesla are building Terafab, a 100 million square foot semiconductor plant in Grimes County, Texas, per Fortune. Elon Musk said it will combine logic chips, memory and the packaging that assembles them under one roof. Texas Gov. Greg Abbott said the first phase amounts to more than $16.8 billion in capital investment and will create 3,000 jobs. The plant is designed to produce more than 1 terawatt of compute a year, the level both companies expect their combined demand to reach. Musk said the site is meant to make AI chips at scale for use on Earth and in space, supporting a plan that also envisions as many as 1 million Tesla Optimus robots.

Read at Fortune ↗

Anthropic buys 191 megawatts of Texas data center capacity from a bitcoin miner

Anthropic signed a $9.1 billion, 20 year agreement with bitcoin mining company Riot Platforms for 191 megawatts of computing capacity at Riot's Rockdale campus in Texas, per Bloomberg. The 191 megawatts is roughly the electricity 143,000 homes draw at any given time. The contract runs through June 2048, with two five year extension options that would lift its potential value to $16.1 billion. Riot's shares rose about 25% to $24.40 in after-hours trading once the Anthropic connection became known.

Read at Bloomberg ↗ • Read at Livemint ↗

Anthropic pitches investors ahead of a September IPO

Anthropic is meeting with potential investors to shore up confidence ahead of an initial public offering, per The Wall Street Journal. The company, valued at $965 billion after a funding round in May, has fielded investor questions about the popularity of cheaper AI systems made in China, its tensions with the Trump administration and a growing backlash to data center construction across the U.S. It is working with Morgan Stanley, Goldman Sachs and JPMorgan Chase on the offering, has filed confidentially and is planning a debut in September or early October. Listing then would put Anthropic on the public markets before OpenAI, which is now looking at 2027 after earlier targeting a fall 2026 debut.

Read at The Wall Street Journal ↗ • Read at Yahoo Finance ↗ • Read at Investing.com ↗

Nature study puts AI's oil and gas emissions above its renewable savings

A peer reviewed study published in a Nature journal estimates that AI's role in raising oil and natural gas output will more than cancel out the emissions it helps avoid by accelerating renewable energy, per Axios. The researchers put the net addition at roughly 1% to 5% of the global energy sector's 2024 emissions, or three to 13 times the International Energy Agency's estimate of current data center emissions. Holly Alpine and Will Alpine, former Microsoft employees who founded the nonprofit Enabled Emissions Campaign in 2024, co-led the study with researchers from Purdue University and an independent researcher. The American Petroleum Institute, the oil industry's main trade group, disputed the premise, saying its members are producing more energy while cutting emissions through better technology and stronger operational practices.

Read at Axios ↗

VI.Global & Geopolitics

India's IT ministry summons Meta and demands an AI content compliance roadmap

India's Ministry of Electronics and Information Technology told Meta to retool its algorithms and hand over a compliance roadmap for curbing deepfakes, propaganda and other AI manipulated content, after the platform temporarily removed a video posted by Prime Minister Narendra Modi, per Biometric Update. At a meeting in New Delhi convened by government summons, Secretary S. Krishnan warned Meta that "compliance cannot be optional," and officials said that if platforms fail to act the government will "invoke blocking powers, court directions and police action." Section 69A of India's Information Technology Act of 2000 lets the central government block public access to online information to protect sovereignty, security and public order, and applies to entire platforms rather than only individual posts. The ministry flagged three failures at Meta, among them missing the three hour takedown window set by this year's amendment to India's rules for online platforms, which carries loss of legal immunity and possible criminal liability.

Read at Biometric Update ↗