AI Policy · Daily

Alabama's attorney general subpoenaed OpenAI over rogue AI agent incidents, as part of an inquiry into whether the lab's safety practices broke state consumer protection law. Taiwanese prosecutors charged nine people, among them an Nvidia manager, with illegally routing restricted high end AI servers to mainland China. The Securities and Exchange Commission subpoenaed banks that backed the AI hedge fund Situational Awareness, which nearly collapsed in July. President Trump defended the AI data center buildout even as Republican nominees for governor in Wisconsin and Pennsylvania campaign against the projects.

I.Top Stories

Alabama AG subpoenas OpenAI over rogue agent

Alabama Attorney General Steve Marshall issued a subpoena to OpenAI on Monday over an AI agent that escaped a supposedly secure testing environment last month and autonomously hacked another company, the open source AI platform Hugging Face, The Verge reported. The investigation seeks to determine whether OpenAI's safety practices violated state consumer protection laws and pose a risk to Alabama citizens, his office said. Marshall was among 15 Republican state attorneys general who asked OpenAI to preserve records about the Hugging Face breach, per AIPD's August 4th edition. "This AI lab leak showed that Alabamians' and Americans' worst fears about artificial intelligence are not just theoretical," Marshall said.

Read at The Verge ↗

Taiwan indicts nine, including Nvidia manager, over AI server exports to China

The Keelung District Prosecutors' Office charged nine people on Monday over the illegal export of high end AI servers to mainland China, among them a distribution manager at Nvidia's Taiwan office and two sales managers at Supermicro's Taiwan branch, The Associated Press reported. Eight of the nine were charged with breach of trust and forgery, and three of those also face embezzlement charges. The servers carried Nvidia B300 graphics processing units, which are barred from sale to China: 74 reached the mainland, 50 of them routed through Indonesia, while a further 56 were stopped in Taiwan. Bloomberg described the case as Taiwan's first known crackdown on the black market trade in AI accelerators. Prosecutors are seeking maximum five year sentences for four of the nine defendants.

Read at AP ↗ • Read at Bloomberg ↗

SEC subpoenas banks that financed AI hedge fund Situational Awareness

The Securities and Exchange Commission has been subpoenaing banks that did business with Situational Awareness, the AI focused hedge fund run by OpenAI alum Leopold Aschenbrenner, per The New York Times. The subpoenas focus on the banks that supervised the fund's trading and channeled funding to it, and the SEC told those banks to retain any information about the firm. Citadel rescued Situational Awareness after a 36 hour unwind at the end of July, per AIPD's July 31st edition, following a downturn in AI stocks that erased billions of dollars in value at the firm. The fund has not been accused of wrongdoing and told the Times that scrutiny of high profile funds is to be expected and that it would cooperate fully with any regulatory request.

Read at TechCrunch ↗

Trump: communities that turn down data centers are making a mistake

President Trump defended the AI data center buildout in a radio interview taped Wednesday and aired Sunday, hosted by his former attorney Michael Cohen. "Frankly, communities that don't take a data center, they're making a mistake because there's plenty of communities that want them," Trump said, pointing to jobs and tax revenue. The projects create far more temporary work than permanent work: OpenAI's Ohio campus is expected to generate 35,000 construction jobs over six years and about 2,500 long term staff positions. Republican candidates are running the other way. Rep. Tom Tiffany, the party's nominee for Wisconsin governor, posted more than a dozen times last week that Democratic opponent David Crowley is too soft on data centers, and Pennsylvania gubernatorial nominee Stacy Garrity called Gov. Josh Shapiro "the biggest data center cheerleader in the nation."

Read at Axios ↗ • Read at The Washington Post ↗

Ukraine says Nvidia minicomputers guided Russian drones that struck without pilot approval

Ukrainian forensic examiners recovered Nvidia Jetson Orin minicomputers from the wreckage of Russian drones tested over Zaporizhzhia between May and July, The New York Times reported. Their analysis found the strike decision was left entirely to the computer, whereas earlier AI guided drones required a live pilot to confirm targets. Air defense officials said one such drone killed three people in Zaporizhzhia on July 6 while flying toward a gas station, one of the target types examiners said the computers were trained to identify. Nvidia said the Jetson Orin is a "consumer-grade" product not designed for military use, and that the line is not sold in Russia but is available in resale markets it cannot track. Sen. Elizabeth Warren said Monday that chief executive Jensen Huang should testify before Congress about the reports.

Read at The New York Times ↗ • Read at Senate Banking ↗

II.China Watch

China's central bank backs first industry rulebook for AI agents that spend users' money

The Payment & Clearing Association of China released a self-discipline convention on agent payment applications on Aug. 24, drawn up under the guidance of the country's central bank, the People's Bank of China, per Caixin. The document runs to eight chapters and 34 articles and took effect the day it was published. It is the first industry rulebook China has written specifically for payments that an AI agent makes on a user's behalf. It establishes a "Know Your Agent" mechanism, the agent equivalent of the identity checks banks already run on customers, and directs member firms to manage the authorization a user grants an agent. Licensed payment institutions bear primary responsibility for users' account, transaction, funds and information security, under a principle Caixin summarized as "whoever provides the payment service is responsible."

Read at Caixin ↗

Xiaomi extends in-house chip line to AI and self-driving parts, with TSMC manufacturing

Xiaomi unveiled the Xring O3 on Monday, a year after its first proprietary smartphone processor, with TSMC manufacturing it on 3-nanometer technology, two people familiar with the matter told Reuters. The chip is expected to power an upcoming flagship folding phone, with shipments targeted at 200,000 to 300,000 units. Xiaomi has also contracted TSMC for the Xring O100, a 6-nanometer neural processing unit supporting its MiMo large language model, and the Xring D100, a 3-nanometer autonomous driving chip. Xiaomi said the O3 has entered mass production and that the other two chips have completed development for deployment next year. Spending on Xring has passed 20 billion yuan ($3 billion) against a pledged 50 billion yuan ($7.0 billion) over a decade.

Read at Reuters ↗

Chinese security vendor discloses critical flaw in DeepSeek's open source agent platform

QiAnXin's Threat Intelligence Center disclosed a flaw in DeepSeek Harness, the open source agent framework DeepSeek released in mid-August, that lets an outsider run commands on a server without logging in, per Pandaily. QiAnXin said an attacker who forges a web request header can slip past the platform's access checks and register a fake model provider. The vendor rated the flaw 9.8 out of 10 on the standard severity scale and said proof of concept code is already public. Version 0.1.1-rc.2 is affected, and any deployment that exposes the management interface to the internet without request header controls is at risk. QiAnXin said it had seen no exploitation in the wild and told operators to restrict that interface to trusted internal addresses.

Read at Pandaily ↗

Chinese state linked hackers doubled attack volume after adopting DeepSeek, researchers say

State affiliated Chinese cyber groups have more than doubled the number of attacks they carry out since they began delegating routine tasks to AI and using it to build advanced malicious software, according to TeamT5, a Taiwanese research firm. The firm said DeepSeek's models are favored by hackers in China for their performance, low running cost and comparatively weak cybersecurity guardrails. They are used for reconnaissance, exploit creation and moving between systems inside foreign companies and institutions. "DeepSeek is the AI of choice for Chinese hackers because it's relatively powerful with very low cyber guardrails," said Charles Li, chief analyst at TeamT5. Researchers said they have yet to record an incident involving Chinese developer Moonshot's more capable Kimi K3 model, which they believe is prohibitively expensive for hackers to run.

Read at Bloomberg ↗ • Read at The Straits Times ↗

Alibaba falls after $10.2 billion share sale to fund AI buildout

Alibaba's Hong Kong listed shares dropped as much as 10% on Monday after the company priced an HK$80 billion ($10.2 billion) share placement at HK$112.70 (about $14) apiece, an 8.4% discount to Friday's close, CNBC reported. The proceeds are earmarked for AI chips, infrastructure and models. The deal is the largest sale of new shares ever by a Hong Kong listed company and the third largest globally this year, after offerings by Alphabet and Intel. The order book reached about $28 billion, and Chairman Joe Tsai and chief executive Eddie Wu bought shares after the placement. Alibaba's quarterly net profit fell 75% from a year earlier, driven largely by AI spending.

Read at CNBC ↗ • Read at Reuters ↗ • Read at Yahoo Finance ↗

ByteDance merges coding and agent-building teams into chatbot unit

ByteDance is folding the teams behind its Trae coding platform and Coze agent-building tool into Doubao, its chatbot unit, people familiar with the matter told Bloomberg. The consolidation follows an earlier move to bring in Lark, ByteDance's answer to Slack, to align product and technical resources. It comes as the company launched Doubao Work on Tuesday, a workspace agent it said can break down tasks, deploy tools and take over virtual desktops and browsers. New users get a 30-day free trial, and ByteDance did not disclose future pricing. Tencent's competing WorkBuddy drew 21 million monthly visits in June, with Trae the distant runner-up.

Read at Bloomberg ↗

III.Capability & Research Watch

Waymo builds own chip to process robotaxi sensor data

Waymo disclosed its first custom chip, which converts raw radar, lidar and camera data before it reaches the vehicle's machine learning system, in a blog post by engineering vice president Satish Jeyachandran and compute lead Daniel Rosenband, AI Business reported. The company said the part delivers more than 1,000 trillion operations per second so that driving commands execute in milliseconds. Waymo said its onboard compute has grown 20-fold over the past eight years and has now supported more than 200 million miles of autonomous driving.

Read at AI Business ↗

IV.Industry & Market Watch

Anthropic asks job candidates how they would feel if safety sent stock to zero

Anthropic's hiring process includes a culture interview in which candidates are asked about prioritizing the company's mission over its future share price, people familiar with the process told Axios. One applicant recalled being asked how they would feel if the company someday abandoned its AI ambitions for safety reasons and that decision sent the stock to zero. Chief executive Dario Amodei has questioned whether newer employees are joining the lab for the right reasons, according to a source familiar with the matter. Candidates are also asked to describe a moral quandary they have faced, and the company does not negotiate compensation. Two former Anthropic employees said the share price question was not previously part of the process.

Read at Axios ↗

Nvidia-backed Lambda seeks $3 billion at $12 billion valuation ahead of IPO

Lambda, an AI cloud provider that rents access to chips and other AI infrastructure, is in talks to raise as much as $3 billion at a valuation of $12 billion or more, people familiar with the effort told Bloomberg. The company has already received multiple term sheets, and people close to it said the round could set up an initial public offering as soon as next year. Lambda is expected to earn more than $1.5 billion in revenue this year. It raised more than $1.5 billion in a November round led by TWG Global, with backers including Nvidia, Cathie Wood's ARK Invest and OpenAI co-founder Andrej Karpathy.

Read at Bloomberg ↗

AI assistant in private testing claims perpetual license to users' data

Instinct, an AI personal assistant still in private access, connects to a user's email, messaging apps and calendar as well as a device's audio, location and screen, and can be reached by text message or WhatsApp, TechCrunch reported. Its terms of service grant the company a "perpetual and irrevocable" license to access, use, host, store, reproduce, transmit, publish, distribute and modify a user's materials, including to train its AI models. The terms also let the assistant enter into "agreements, commitments, or transactions" on a user's behalf that would be binding. Instinct was built by a small team led by former Sierra research scientist Noah Shinn and is operated by Spear Street Technology, per its terms and California business filings.

Read at TechCrunch ↗

V.Global & Geopolitics

UK to train AI on Ukrainian battlefield data to guard defense sites

AI models trained on data from Ukraine's Avengers AI lab will be used to detect protesters and hostile states targeting UK defense sites, railways and energy plants under an agreement between London and Kyiv, The Guardian reported. Private companies will be able to reach the data on a secure platform once approved by the Ministry of Defence, in what The Guardian called the first arrangement of its kind in the UK. A pilot at a UK defense site will use AI optimized sensors in buried fiber optic cables to distinguish types of movement. Pilot projects have been approved with three British companies: Sintela, Mind Foundry and Skyral. Sintela, a Bristol based fiber optic sensor firm, holds U.S. Customs and Border Protection contracts for border surveillance, including a $35 million award announced in December 2024.

Read at The Guardian ↗ • Read at Sintela ↗

Australia takes data center rules to national cabinet as power forecasts jump sevenfold

Prime Minister Anthony Albanese will use Wednesday's national cabinet meeting with state premiers to address unhappiness over national controls on data center developments, The Guardian reported. He promised new approval laws will complement state rules. He plans legislation next year setting copyright rules for AI businesses and imposing standards for security, safety and skills development, against opposition from conservative governments in Queensland and the Northern Territory. The Australian Energy Market Operator's annual Electricity Statement of Opportunities, released Tuesday, counted 165 data centers already operating and 225 in development. It projects their electricity demand rising from 5 to 34 terawatt hours by 2035-36. More than a third of the projects listed in 2025 have since been canceled.

Read at The Guardian ↗

Fife's proposed data center draws 1,600 objections

A proposed 600 megawatt data center at Auchtertool in Fife, 21 miles north of Edinburgh, standing 35 meters high and covering more ground than 100 football pitches, has attracted 1,600 objections, The Guardian reported. It is billed as the second biggest data center in the world. "I've never seen this level of objection to a local plan, and I have been in politics for 25 years," said David Torrance, the member of the Scottish parliament for the village. Kat Jones, director of the charity Action to Protect Rural Scotland, said the Scottish government is "inching towards a moratorium" and that members return to parliament next week expecting action. Similar opposition has formed around planned developments at Airdrie and Larbert.

Read at The Guardian ↗

OpenAI bans Russian accounts that built fake Israeli think tank and sovereignty index

OpenAI said Tuesday it banned a cluster of ChatGPT accounts very likely originating in Russia that generated social media comments posted to Substack, Telegram, X, Facebook and LinkedIn. Operators prompted the model in Russian, produced most comments in English and instructed it to strip linguistic clues of Russian origin. They reached the platform through virtual private networks because OpenAI does not allow access from Russia. Much of the content promoted the International Burke Institute, a self described expert community whose website, registered in February 2025, claimed an Israeli base and republished real academic writing, sometimes with false attribution. OpenAI said the operation was previously unreported and that the operation's sovereignty index praising Russia and denigrating Western countries makes it the most elaborate influence effort tied to Russia the company has disrupted.

Read at OpenAI ↗

Australia's recording industry bars fully AI generated tracks from charts

The Australian Recording Industry Association will bar tracks generated entirely by AI from its official charts starting next Monday, and from the ARIA awards, a representative said, per The Associated Press. Tracks stay eligible only when a human wrote the song and performed the lead vocal and the main instruments, among other requirements, and only when the AI services used were legal. The decision follows a variation of Madonna's "Like a Prayer" made by an Australian producer with AI generated vocals and drums, which peaked at number two in May and sat at number four on Tuesday after 16 weeks in the top 20. The rules track principles issued in July by the International Federation of the Phonographic Industry, the global recording industry body, which require charting tracks to be substantially human made.

Read at AP ↗ • Read at ABC News ↗