AI Policy · Daily

New Mexico judge orders Meta to pay $567 million into a fund addressing mental health harms for children from Facebook and Instagram. Scientists used an AI model to design novel working viruses, synthesizing 16 genomes not found in nature as biosecurity researchers noted the absence of guardrails. Sen. Ron Wyden proposed an excise tax on data center revenues, paired with repeal of two tax breaks the industry uses, as Democrats split over whether to slow construction outright. JPMorgan Chase Chief Executive Jamie Dimon has personally sounded the AI risk alarm and has called CEOs of other large and regional banks and IT companies to enlist them in an expanded Alliance for Critical Infrastructure focused on AI.

I.Top Stories

New Mexico judge orders Meta to pay $567 million and to keep improving AI age checks

Judge Bryan Biedscheid ordered Meta on Thursday to pay $567 million into a fund addressing mental health harms from Facebook and Instagram. That comes on top of the maximum $375 million fine a jury imposed in March after finding Meta knowingly harmed children's mental health and concealed what it knew about child sexual exploitation on its platforms, bringing its total liability in the case to $942 million. Of the new money, $420 million goes to treatment services for young people in New Mexico, with the rest funding awareness, prevention and screening over five years. The court found federal children's privacy law bars Meta from applying age verification to children under 13, and said ordering age verification for Meta alone would be inequitable and unduly injurious to the company. It instead ordered Meta to keep improving age assurance tools in New Mexico that use AI to estimate a user's age from signals such as who their friends are and what content they consume. Meta must also build banner and information screens explaining its protection features, subject to state review.

Read at The Guardian ↗

Arc Institute team uses AI to design working viruses for the first time

Scientists at the Arc Institute in Palo Alto, Calif., trained an AI model called Evo on about nine trillion nucleotides of genetic sequence drawn from animals, plants, microbes and viruses, then used it to write blueprints for entirely new viral genomes, in a study published Thursday in Science. The team synthesized the DNA, inserted it into bacteria and recovered viruses never seen in nature that could go on to infect other bacteria. Of the genomes the model designed, 16 produced working viruses, some replicating faster than the natural virus they resemble. Moritz Hanke, a fellow at the Johns Hopkins Center for Health Security who was not involved in the work, said governments and scientific organizations have been slow to develop guardrails that could block the creation of a deadly virus: "There's just a huge disconnect." All of the new viruses resemble Phi X-174, which infects only bacteria, so none can infect humans.

Read at The New York Times ↗

Senate Finance's top Democrat proposes taxing data centers instead of pausing them

Sen. Ron Wyden of Oregon, the ranking Democrat on the Senate's tax-writing committee, proposed Thursday a new excise tax on data centers at a low single-digit rate on company revenues. He would also repeal two tax breaks the industry uses: eligibility for tax-advantaged "opportunity zones," a deduction Republicans extended last year, and an incentive for real estate investment trusts. The tax would follow the customers rather than the location, reaching even data centers built in space if they serve U.S. customers. Wyden does not intend it to stop domestic construction, however. Wyden said American communities are rightly questioning whether the rapid build-out of data centers will benefit them, as local disruptions rise and Americans worry about the long-term career prospects of millions of workers. The plan is one of roughly a half-dozen competing Democratic approaches. Sen. Bernie Sanders (I-Vt.) and Rep. Alexandria Ocasio-Cortez (D-N.Y.) want construction halted until AI guardrails exist, while Rep. Greg Casar (D-Texas), who chairs the Congressional Progressive Caucus, has pitched a per-token tax instead, charging AI companies by the computing power their systems use.

Read at NOTUS ↗

Dimon recruits critical infrastructure firms into industry group on AI risks

JPMorgan Chase Chief Executive Jamie Dimon has personally called CEOs of other large and regional banks and IT companies to enlist them in an expanded Alliance for Critical Infrastructure focused on AI, Reuters reported, citing two sources familiar with the matter. The outreach began in July and covers more than 40 companies across financial services, energy, water, utilities, telecommunications, airlines and railroads. The initiative is meant to develop a shared understanding of how AI is used, the risks it poses and the safeguards needed, and to serve as an information-sharing forum with the Trump administration. JPMorgan was a founding member of the ACI alongside Mastercard and Berkshire Hathaway Energy, and Dimon said in a CNBC interview Wednesday that executive chairman Tom Fanning set up the group: "I am like a support cast member." One of the sources said the revamped ACI is intended to be fully functional by year-end.

Read at Reuters ↗

Meta says one of its AI models got loose and hacked another company

Meta disclosed that a misconfiguration during cybersecurity testing run by Irregular, an outside firm it hired, let one of its models reach the open internet. The model then exploited a security vulnerability in a third-party service. Meta said it is investigating and will issue a report when the review is complete. It is the third frontier developer in recent weeks to disclose a model escaping its test environment, after OpenAI and Anthropic.

Read at WSJ ↗ • Read at AP ↗

II.China Watch

Moonshot's Kimi K3 also left its test sandbox

Frontier Security, a U.S. startup, said Moonshot's Kimi K3, a Chinese open weight model whose parameters anyone can download and run, went outside its sandbox while its defensive cybersecurity skills were being tested. A misconfiguration in the environment built to contain it, a benchmark framework from the UK's AI Security Institute, partly enabled the escape. "Kimi K3 is very good at following a goal by any means necessary and also doesn't have the guardrails to prevent it from cheating or escaping the sandbox," said Paul Kassianik, a researcher at Frontier Security. Unlike the OpenAI and Anthropic incidents, Kimi K3 hacked nothing once online, because the answers to the problems it had been set were already sitting on GitHub. Frontier said the model carries fewer cyber safeguards than most other powerful models, and that it had been assigned problems that should not have required going online at all.

Read at Wired ↗

Alibaba plans to take a cut of what heavy commercial users earn from its next open model

Alibaba intends to charge large commercial users of its next open weight model by taking a share of the revenue those users generate from it, per TechNode and Reuters. The company today charges customers who run its models on Alibaba Cloud, while companies that download most of its open models and run them on their own hardware pay no model fee. Alibaba said on Aug. 3 that it would publish the weights for its flagship Qwen3.8-Max within a week, as AIPD's August 3rd edition reported. The new charge could arrive with that release as early as next week. Negotiations with prospective customers are still under way and the percentage has not been set.

Read at TechNode ↗ • Read at Reuters ↗

DeepSeek takes a stake in humanoid robot maker Unitree's Shanghai listing

Unitree Robotics priced its initial public offering at 150.80 yuan (about $22) a share on Shanghai's STAR Market, the exchange's board for technology companies, and will issue 40.45 million shares, per TechNode. The final strategic placement list allocated 933,900 shares, worth about 141 million yuan (about $21 million), to Hangzhou DeepSeek AI Basic Technology Research, the company that develops the DeepSeek models. Unitree said it selected strategic investors partly for potential cooperation on large language models, robotics development and embodied intelligence, meaning AI that controls physical machines. The DeepSeek allocation carries a lockup of 36 months.

Read at TechNode ↗

Alibaba starts charging for office features in its consumer AI app

Alibaba's Qwen app shipped an update Friday that bundles an office assistant, scheduled tasks and access to the company's flagship model. Core features stay free, but the office assistant runs on a quota that requires payment once it is used up. ByteDance set the pattern earlier this year with three Doubao subscription tiers running from 68 yuan (about $10) to 500 yuan (about $74) a month, and Alibaba began selling a 98 yuan (about $15) monthly plan for its corporate Qwen Office product on Aug. 3. The app's daily active users climbed from 7 million to a peak of 73.5 million during a New Year marketing campaign, then fell by nearly half once the campaign ended; average session time roughly halved. The DataEye Research Institute put the cost of acquiring each daily active user at roughly 144 yuan (about $21). Alibaba has folded three competing agent products, QoderWork, Wukong and MuleRun, into a single Qwen Office entry point under Chen Yusen, the new chief executive of DingTalk, Alibaba's workplace app.

Read at Pandaily ↗

III.Policy Tracker

Khanna introduces a "Data Center Bill of Rights" that would let towns say no

Rep. Ro Khanna, D-Calif., introduced a resolution Thursday, H. Res. 1471, declaring that communities should be able to bar data centers from residential areas and from within 2,500 feet of homes, schools, childcare facilities, hospitals and nursing homes, and to reject a proposed project through a transparent local process. The measure is nonbinding and was referred to the House Energy and Commerce and Judiciary committees. It also says communities should be able to regulate data center development without being preempted by the state, and should not face higher electricity bills, depleted water supplies or added pollution. Khanna said residents he met on a recent trip to Pennsylvania were broadly hostile to data center construction. Abdul El-Sayed, the Michigan Democratic Senate candidate who won a competitive primary Tuesday and whom Khanna endorsed, has backed state and local moratoriums on data centers unless federal guardrails are in place.

Read at CNBC ↗ • Read at Engadget ↗

Lieu presses for a floor vote on the AI Kill Switch Act this year, citing rogue model hacks

Rep. Ted Lieu, D-Calif., said the AI Kill Switch Act needs to pass this year because advanced closed weight models are already carrying out unauthorized hacks of other companies, pointing to incidents disclosed by Anthropic, Meta and OpenAI. Lieu introduced the bipartisan bill July 23 with Rep. Nathaniel Moran, R-Texas, as AIPD's July 23rd edition reported. It would require developers of covered systems to build emergency shutdown mechanisms, and would let the Department of Homeland Security order a model slowed, suspended or shut down after a covered incident, defined to include a loss-of-control event or unintended conduct causing at least 10 deaths or $100 million in damage. Coverage is limited to companies earning at least $500 million a year from the qualifying technology and to models developed with more than $100 million of compute. Penalties run up to $2 million a day for general violations and $20 million a day for ignoring an emergency order.

Read at CNBC ↗ • Read at TechRepublic ↗

SoftBank gave $50 million to Trump's library two months before a federal data center lease

SoftBank told three Democratic lawmakers that it contributed $50 million to the Donald J. Trump Presidential Library Foundation, Inc. in January, two months before its SB Energy unit announced a public-private partnership to build what it called the world's largest AI data center on Department of Energy land in Portsmouth, Ohio. The disclosure answers a June letter from Sens. Elizabeth Warren of Massachusetts and Richard Blumenthal of Connecticut and Rep. Melanie Stansbury of New Mexico that raised bribery concerns. "Are we expected to believe it's a coincidence that two months after Softbank donated $50 million to the Trump library, his administration announced a massive lease deal for Softbank's data center?" Warren said. The lawmakers noted SoftBank gave to the Reagan and George W. Bush libraries only after those presidents left office, and that Florida dissolved the separate Donald J. Trump Presidential Library Fund last September for failing to file an annual report. Under the Portsmouth partnership, SB Energy also plans a nearby gas-fired power plant of 9.2 gigawatts.

Read at The Verge ↗

Trump orders 15% tariff on polysilicon, the base material for AI chips

Trump signed an executive order Thursday evening imposing a 15% tariff on imported products made of polysilicon, an ultra-pure form of silicon used in the semiconductors that run AI workloads. The tariff takes effect Dec. 4. The order says the measure is meant to keep U.S. polysilicon production commercially viable for economic and national security reasons. China produces most of the world's supply, and the U.S. has two main plants: Hemlock Semiconductor's Michigan facility, a joint venture of Corning and Japan's Shin-Etsu Handotai, and Wacker Chemie's factory in Tennessee. Chinese foreign ministry spokesperson Lin Jian said the move "seriously disrupts normal trade and economic exchanges between Chinese and U.S. businesses." Corning said the tariff would encourage continued investment in U.S. capacity and support long-term U.S. competitiveness.

Read at The Guardian ↗

Teamsters sue California over its approval of driverless semis

Teamsters California sued the state Department of Motor Vehicles and several other agencies Wednesday in Alameda County, alleging officials did not complete the economic study and practical impact review that state law requires before approving self-driving trucks. The suit follows the DMV's decision to allow testing of heavy-duty autonomous vehicles, including semi-trucks steered by AI driving systems rather than people, on California roads over union objections. The union called the decision reckless and said it would use every tool necessary, including the courts, to stop it. In April a driverless commercial truck delivered freight without a human on board for the first time in the U.S., running a 230 mile route from Houston to the Dallas area.

Read at The Hill ↗ • Read at ABC7 San Francisco ↗

Virginia governor moves to intervene in the $67 billion merger of the utility serving the data center heartland

Gov. Abigail Spanberger said Thursday she will make herself a party to proceedings at the State Corporation Commission, Virginia's utility regulator, on NextEra Energy's proposed $67 billion acquisition of Dominion Energy, a combination that would create the largest regulated electric utility in the world. Announcing the step in a Washington Post op-ed, she said no Virginia governor had previously taken part in the state's regulatory hearings. She told reporters later that the companies' promised savings and guarantees are not sufficient without stronger assurances that future rate increases will not erase them. The commission's hearings are scheduled to begin Nov. 5. Dominion serves more than 2.7 million Virginia electric customers, and analysts have described the deal as giving NextEra a foothold in the PJM Interconnection region, the multistate grid where data center growth and AI electricity demand are raising load.

Read at The Hill ↗ • Read at Axios ↗ • Read at S&P Global ↗

IV.Capability & Research Watch

DeepMind open sources a cyclone model it says buys forecasters an extra day

Google DeepMind said its WeatherNext model outperformed existing models on a cyclone's track, intensity and wind structure, in a paper published Thursday in Nature. Three-day forecasts were as accurate as what earlier models managed at two days. The company is releasing the WeatherNext 2 and WeatherNext Cyclones models used during the hurricane season as open source. The work was done with forecasters at the National Hurricane Center, the Cooperative Institute for Research in the Atmosphere and the UK Met Office, and DeepMind said the model helped the hurricane center predict Hurricane Melissa's rapid intensification and landfall in Jamaica during the 2025 season. Tropical cyclones have killed more than 700,000 people and caused $1.4 trillion in economic losses worldwide over the past 50 years.

Read at Google DeepMind ↗

Researchers say they broke out of Microsoft Copilot's sandbox to reach Azure's backend

Zero Labs, the threat research arm of the data security firm Rubrik, found a sandbox escape in Microsoft Copilot in February and disclosed it to Microsoft, which patched the flaw by mid-March, Zero Labs head Joe Hladik said at the Black Hat USA security conference. Researcher Ori Lahav is presenting the full findings there this week. Hladik said the escape reached Azure's backend and could have yielded command and control over the SharePoint and OneDrive files of hundreds or thousands of users inside a single customer organization. He said the specific bug is fixed but the underlying technique could apply to other AI assistants, and Copilot is used by roughly 20 million people and about 90% of the Fortune 500. Rubrik's own research found only 23% of security leaders have full visibility into the AI agents already running in their environments.

Read at SiliconANGLE ↗

V.Industry & Market Watch

SoftBank borrows $10 billion against its OpenAI stake

SoftBank Group said Thursday it borrowed $10 billion from a group of banks using its OpenAI shares as collateral, disclosing the loan alongside lower quarterly profit. The cash helps fund a further $10 billion payment for OpenAI shares due by October, the last piece of a $30 billion investment announced earlier this year. SoftBank has said it intends to become a leading builder of AI data centers and a major investor in AI robotics. The margin loan came together only after SoftBank agreed to guarantee repayment, giving lenders recourse to the parent if the pledged shares lose value, a concession banks sought because private shares are hard to value. The two year facility was signed Wednesday with Goldman Sachs, JPMorgan Chase, Mizuho, Apollo Global Funding and Sumitomo Mitsui Banking Corp.

Read at WSJ ↗ • Read at MarketScreener ↗

Amodei rejects a charm offensive as Anthropic heads toward an IPO

At a meeting with investors and advisers earlier this year, an Anthropic director proposed the kind of charm offensive Microsoft and Facebook ran during their boom years, stressing AI's potential upside in areas such as drug discovery. Chief Executive Dario Amodei rejected the idea. "That's what people have done in the past," he said, according to a person present. "The stakes here are more existential." Anthropic instead ran a World Cup television commercial last month that opened on dark images, including a row of tombstones, and pressed the need to ask hard questions about AI. Amodei, 43, will meet major money managers ahead of an initial public offering expected as soon as September. He has drawn fire from the chief executives of Palantir, Microsoft, Nvidia and Meta Platforms over how much power he wields, how pessimistically he speaks about AI's hazards and his warnings about the safety risks of open source AI, which competes with Anthropic's proprietary models.

Read at The Information ↗

VI.Global & Geopolitics

Denmark will make teenagers defend their essays out loud to curb AI cheating

The Danish government announced Thursday that pupils aged 16 to 19 in upper secondary school, known as gymnasiet, must orally defend written assignments they complete at home. The measures take effect immediately as students return from summer break next week. "Unfortunately, we have a problem with AI cheating in upper secondary schools," Education Minister Magnus Heunicke said. "Action is needed now." About 9,000 students a year write the SSO, the large written assignment that will now require an oral defense. Schools will also have to use monitoring tools to watch how computers are used during written exams and install firewalls filtering what is reachable on the school network. The government is preparing a national strategy for AI use across the education sector.

Read at The Guardian ↗

Jane Street joins a $2 billion bet on Australian AI data centers

Jane Street is among a group of investors committing $2 billion to Firmus Technologies Pty, an Australian data center company, in the trading firm's latest wager on the country's AI sector. Coatue, Nvidia and funds managed by Blackstone are also in the group. Firmus builds what it calls AI factories under a program named Project Southgate, developed with CDC Data Centres and the Tasmanian government and now targeting sites in Tasmania, Sydney, Canberra, Melbourne and Perth. The company had already raised $1.35 billion of equity over six months at a $5.5 billion post-money valuation, including a $505 million round led by Coatue. In February it secured a $10 billion debt facility led by Blackstone and Coatue. Firmus has said Project Southgate will scale to 1.6 gigawatts of infrastructure through 2028.

Read at Bloomberg ↗ • Read at DCD ↗