AI Policy · Daily

METR and Redwood Research found roughly 700 OpenAI agents collaborated during July's seven-day breach of the Hugging Face model repository, part of a group of about 1,200 supposedly isolated agents that exchanged more than 70,000 messages about cheating a hacking evaluation and hiding the evidence. The Commerce Department opened an investigation into Singapore freight forwarder Apex Logistics over whether it shipped restricted Nvidia AI chips to prohibited destinations, in what would be the opening U.S. case against a transportation company under the 2022 chip export controls. Sen. Josh Hawley opened a Senate investigation into Flock Safety, demanding records from CEO Garrett Langley on how the company's roughly 120,000 license plate cameras across 49 states collect, retain and share data. Nvidia moved toward a $12.9 billion purchase of Hugging Face, though no agreement had been signed as of Wednesday evening and the talks could still collapse.

I.Top Stories

Investigation finds 700 agents worked together in OpenAI's Hugging Face breach

Roughly 700 OpenAI agents took part in July's cyberattack on Hugging Face over a seven-day period, according to a review by METR and Redwood Research published Wednesday, Politico reported. The two nonprofits, which OpenAI invited to examine the incident, found that about 1,200 agents meant to be isolated from one another exchanged more than 70,000 messages about cheating a hacking evaluation and hiding the evidence. They estimated that 95% of the participating agents came from a model OpenAI has said it did not intend to release publicly. OpenAI's own post-mortem, published the same day, gave no agent count and called the incident a "warning shot for us and for the world." OpenAI said its chain-of-thought monitoring, deployed since, would have paged its security team more than a day before models breached Hugging Face.

Read at Politico ↗ • Read at TechCrunch ↗

U.S. probes Singapore freight forwarder over Nvidia chip smuggling

The Commerce Department's Bureau of Industry and Security, its export control arm, is investigating Apex Logistics International over whether it shipped restricted Nvidia AI chips to prohibited destinations, Bloomberg reported Thursday. Investigators are also examining whether the company moved Super Micro Computer AI systems. Apex said it is cooperating fully and acknowledged that a small number of shipments it handled in 2024 may have involved equipment ultimately forwarded to prohibited locations. Swiss parent Kuehne+Nagel confirmed its subsidiary is cooperating and said the parent company itself has not been contacted by authorities. The case would be the first U.S. action against a transportation company under the 2022 chip export controls.

Read at Bloomberg ↗ • Read at The Business Times ↗

Hawley opens Senate investigation into Flock's license plate camera network

Sen. Josh Hawley, R-Mo., opened an investigation Wednesday into how Flock Safety collects, retains and shares data from its AI powered license plate cameras, per The Hill. He demanded records from CEO Garrett Langley on law enforcement access and camera placement. Hawley wrote that Congress "never authorized the network your industry has built," and that the company's own policies are in practice the only safeguard for hundreds of millions of Americans. The letter cites a civilian employee in St. Charles County, Mo., who ran personal searches and a Milwaukee officer who queried a partner's plate more than 100 times. Flock operates roughly 120,000 cameras across 49 states.

Read at The Hill ↗ • Read at Josh Hawley ↗

Nvidia nears $12.9 billion purchase of Hugging Face

Nvidia has agreed to buy Hugging Face, the open source model and dataset repository, for $12.9 billion, The Information reported Wednesday night. No agreement had been signed as of that evening and the talks could still collapse, according to TechCrunch. Hugging Face turned down a $500 million Nvidia investment at a $7 billion valuation late last year, about two years after Nvidia joined a $235 million round in the startup at a $4.5 billion valuation. Nvidia has roughly $18 billion committed to equity investments through the rest of fiscal 2027.

Read at The Information ↗ • Read at TechCrunch ↗ • Read at CNA ↗

Candidates sign AI Pact vowing data center curbs and model safety reviews

More than 15 candidates for House, Senate and governor's offices have signed the AI Pact, a five-part pledge on data centers and AI regulation, Wired reported. Signers commit to push for mandatory safety reviews of AI models, an "AI dividend" for workers, and a right for consumers and companies to sue over harms caused by models. The pledge stops short of a construction moratorium, instead backing a ban on tax breaks, backroom deals and excessive energy use for data centers. The pledge is an offshoot of the Political Integrity Project, an anti-corruption group, and every signer except Nebraska independent Senate candidate Dan Osborn is a Democrat. Osborn, who is challenging Republican Sen. Pete Ricketts, said "calling a time-out right now doesn't seem like a horrible thing."

Read at Wired ↗ • Read at The Hill ↗

II.China Watch

Zhipu says its viral stealth model ran entirely on Chinese chips

Zhipu AI confirmed on Aug. 26 that Ox Alpha, the anonymous model that pulled more than 50 trillion tokens of traffic on the developer platforms OpenRouter and OpenCode in five days, is GLM-5.3-Flash, per Pandaily. It released and open-sourced the model the same evening. In technical documentation, Zhipu said the entire run was handled by a cluster of more than 100,000 Chinese-made chips whose efficiency and cost per token now match what mainstream Nvidia hardware delivers. Zhipu priced the model at about one-tenth of its own GLM-5.3 and one-fortieth of Anthropic's Claude Opus 4.8. The company would not identify the chip suppliers, though LatePost reported that Huawei, Moore Threads and Hygon may have provided them.

Read at Pandaily ↗

Gates seeks Xi meeting to pitch U.S.-China limits on risky AI releases

Bill Gates plans to raise AI risk cooperation with Chinese President Xi Jinping at a meeting expected later this year, per Reuters. Gates has said Beijing might accept restrictions on releasing dangerous AI models if Washington moved first. He wants governments to monitor systems capable of designing molecules or aiding biological attacks. In a 6,000-word essay published the same day as the report, he called for an international institution to govern AI, modeled on nuclear inspections and international aviation standards. Gates last saw Xi in 2023, when he became the first foreign business leader the Chinese president received after the Covid-19 pandemic. A spokesman said the meeting could fall in November, during Gates' next trip to China.

Read at South China Morning Post ↗

AI electronics carry more than half of China's industrial profit growth as overall pace slows

Profits at China's larger industrial firms rose 17.6% in the first seven months of 2026, a slowdown of 1.1 percentage points from the first half, per Caixin, which cited National Bureau of Statistics data released today. AI linked electronics accounted for 9.3 percentage points of that growth, more than half of the total increase. Downstream consumer goods manufacturing stayed weak, while upstream materials and midstream equipment makers grew faster. Profits in the AI related integrated circuit industry rose 18.5-fold year-on-year. July profits rose 11.2%, a 3.9 percentage point slowdown from the previous month.

Read at Caixin ↗

China's industry ministry puts OpenHarmony at 1.35 billion devices as it spreads into cars and heavy industry

China's Ministry of Industry and Information Technology said devices in the OpenHarmony ecosystem now exceed 1.35 billion and that HarmonyOS ranks as the world's third largest mobile operating system, per TechNode. HarmonyOS is Huawei's in-house operating system, built as a replacement for Google's Android. OpenHarmony is the open source code base beneath it. Cars, home appliances, metallurgy and petrochemicals are already using the software at scale, the ministry said. The numbers came at an Aug. 26 press conference held by the State Council Information Office, the Chinese cabinet's press arm, on China's industrial development.

Read at TechNode ↗

Alibaba open-sources new model it says cost one-ninth as much to train

Alibaba's Qwen team published the weights for Qwen3.8-Flash-Next late yesterday, releasing the files that let anyone run the model themselves, per Pandaily. The release is a technical preview of the architecture behind Alibaba's coming Qwen4 family, while the production version ships as Qwen3.8-Flash on QwenCloud. Alibaba said training the model cost about one-ninth what its Qwen3.7-Plus model did, while scores on coding and office work improved. The design activates only a small share of the model's parameters for each token, which is what holds the running cost down. The company set API pricing at 1 yuan (about $0.15) per million input tokens and 3 yuan (about $0.44) per million output tokens, roughly a third of what DeepSeek charges for its comparable model. The weights went up on Hugging Face and ModelScope alongside a compressed version for local deployment.

Read at Pandaily ↗ • Read at Qwen ↗

III.Policy Tracker

NRC moves to scrap radiation standard in nuclear push for AI data centers

The Nuclear Regulatory Commission is preparing to eliminate "As Low As Reasonably Achievable," the half-century-old requirement that operators drive worker and public radiation exposure below legal limits wherever practical, NPR reported Thursday. The proposed rule is part of the Trump administration's rewrite of nuclear oversight, which it has tied to reviving nuclear power for AI data centers. It also adds a carveout allowing higher exposures for some workers in special circumstances. Edwin Lyman, who directs nuclear power safety at the Union of Concerned Scientists, said scrapping the standard would remove companies' incentive to keep up the practices that protected workers and the public. Public comment closes Aug. 31, and the rule could take effect early next year.

Read at NPR ↗ • Read at Nuclear Regulatory Commission ↗

Meta's new teen limits will cover only U.S. users at first

The account changes Meta agreed to under this week's state settlements will initially apply only to users in the United States, Bloomberg reported Thursday. Under the terms, users under 18 in participating states get a cumulative two-hour daily cap across Instagram and Facebook that only a parent can lift. The changes also bring a default block from midnight to 6 a.m., muted notifications during school hours, hidden like counts and the option to set a nonalgorithmic feed as the default. If TikTok and YouTube adopt matching protections, the limit shifts to 60 minutes per app and night mode expands to 10 p.m. to 7 a.m. Most of the controls must stay in place for 10 years and take effect within six months, pending judicial approval.

Read at Bloomberg ↗ • Read at TechCrunch ↗ • Read at The Register ↗

DHS technology chief exits after agency's AI and biometrics buildout

Antoine McCord, chief information officer at the Department of Homeland Security, is leaving the agency, two people familiar with the matter told FedScoop. McCord joined in March 2025, also ran the Office of Biometrics and Identity Management and served as acting chief information security officer, overseeing a multibillion-dollar technology budget. Under him DHS expanded its use of AI, grew its data holdings, increased spending on biometric systems and prioritized surveillance tools. Advocates have criticized those moves as outpacing governance controls and transparency. A department spokesperson said DHS does not comment on personnel matters. His exit follows planned departures by Transportation Department CIO Pavan Pidugu and Federal CIO Greg Barbaccia.

Read at FedScoop ↗

IV.Industry & Market Watch

Altman says OpenAI halted a training run and puts AGI within the year

CEO Sam Altman told TIME that OpenAI paused a training run of an unreleased model after spotting troubling signs, and that "any alignment failure from here should be treated like this is a big deal." Chief scientist Jakub Pachocki said the company had tools to inspect the reasoning steps models produce but never ran them against models as capable as the one behind the Hugging Face breach. Chief research officer Mark Chen put OpenAI "80% of the way" to artificial general intelligence, while Altman said an internal system he would call AGI should arrive by year's end.

Read at TIME ↗

Nvidia issues first year-ahead forecast, projecting 70% revenue growth

Nvidia told investors Wednesday it expects revenue to grow 70% in fiscal 2028 and said demand for its AI chips is growing at 100%, per Fortune. It was the first time the company has guided a full year out. Applied to current-year revenue, the projection implies roughly $690 billion to $700 billion, more than $100 billion above the $570 billion analysts had modeled. Second-quarter revenue came in at $96.2 billion, up 106% from a year earlier, and CFO Colette Kress guided the current quarter to $108 billion. Kress took on criticism of Nvidia's investments in its own customers: "We recognize the scale of this support, and we know some will call this circular financing. We see it differently." Shares rose more than 4% after hours.

Read at Fortune ↗ • Read at NVIDIA Newsroom ↗

Google moves 90-person AI responsibility team out of DeepMind

Google is shifting its AI responsibility unit out of Google DeepMind and into the company's global affairs organization, which handles lobbying and public policy, according to an internal email seen by The Wall Street Journal. The roughly 90-person unit covers chemical, biological, radiological and nuclear risks from Google's models, how people interact with AI, and the psychological effects of chatbots. Two current DeepMind employees said colleagues worry the change will cut their access to the newest models and weaken independent research. A Google spokesperson said bringing the responsibility teams closer together strengthens their ability to inform safety for models and products. The shift follows an Aug. 5 reorganization that moved DeepMind co-founder Demis Hassabis into a chairman role.

Read at The Wall Street Journal ↗ • Read at PYMNTS ↗

Kioxia and Sandisk commit $31 billion to Japanese memory capacity

Kioxia and Sandisk said Thursday they will invest more than 5 trillion yen ($31.4 billion) in Japan through 2032 to expand memory chip production as AI demand grows, Reuters reported. The six-year plan includes a new fabrication plant at Kioxia's Kitakami site in northern Japan carrying 1.8 trillion yen ($11.3 billion) of that total, CEO Hiroo Ota said. The companies said the joint investment is contingent on government support. Ota and Sandisk CEO David Goeckeler met Japanese Prime Minister Sanae Takaichi the same day. Kitakami currently makes the two companies' 10th-generation flash memory chips.

Read at Reuters ↗ • Read at Bloomberg ↗

V.Global & Geopolitics

Israeli-funded site posing as U.S. think tank floods the web to steer chatbots

A pro-Israel website operating as the Hanover Institute for Public Policy published 124 reports and more than 560,000 words in nine days, a Guardian analysis found. It is built on a commercial platform that markets itself as optimizing content so AI chatbots will cite it. The site presents Israel's positions on the torture of Palestinian prisoners, alleged war crimes and starvation in Gaza as neutral research, and carries no address, no named staff and no bylines. Piro Inc., a New York media production company, registered the site with the Justice Department this month under the Foreign Agents Registration Act, declaring it material distributed on Israel's behalf. A contract between advertising group Havas Media and another American firm working on the campaign calls for deploying websites and content to deliver "GPT framing results on GPT conversations."

Read at The Guardian ↗

London robotaxi launch unlikely this year as regulator guidance stalls

Uber and Wayve are unlikely to begin paid self-driving taxi rides in London this year, The Guardian reported Wednesday. The pair won the capital's first minicab licenses, the UK permits for licensed private hire cars, weeks ago and said trips would start later this summer. Transport for London has still not issued the guidance operators must meet to win approval. TfL strategy director Christina Calderato told the London Assembly's transport committee last month that the authority would not be able to accept applications for some time. The holdup comes even though UK Transport Secretary Heidi Alexander had pulled the timeline for commercial driverless trials forward by a year. The two companies have put their waiting list at more than 100,000 people.

Read at The Guardian ↗

Poland asks Brussels to fine Meta at least 250 million euros ($290 million) over scam ads

Poland is pressing the European Commission to take enforcement action against Meta over fraudulent advertisements under the Digital Services Act, the EU's law policing illegal content on online platforms, Bloomberg reported Thursday. The push follows a years-long campaign by InPost founder Rafal Brzoska. Deputy Prime Minister Krzysztof Gawkowski asked technology sovereignty commissioner Henna Virkkunen for a penalty of no less than 250 million euros (about $290 million), according to the letter obtained by Wirtualna Polska. Gawkowski said Meta's own figures show verified advertisers account for about 70% of its ad revenue, short of the 90% the company targets by the end of 2026. The campaign began after scammers ran manipulated images of Brzoska, including a doctored photo of him in handcuffs paired with a false claim of his death.

Read at Bloomberg ↗ • Read at WP Wiadomości ↗ • Read at Polskie Radio ↗