AI Policy · Daily

OpenAI's Sam Altman canvassed the White House and Congress this week ahead of the administration's Aug. 1 deadline for its framework reviewing frontier AI models, telling one senator he backs public equity stakes in AI companies but not at the level Democrats have proposed. METR, Redwood launch independent review of OpenAI Hugging Face incident. The developers of HAWK withdrew the encryption scheme from NIST's post-quantum standards contest after an Anthropic model cut its effective key strength in half, though the tested versions were deliberately weakened rather than deployed systems. China's Commerce Ministry promised retaliation over the FCC's import ban on foreign-made humanoid robots, robot dogs and power inverters, weeks before Xi Jinping is expected in the U.S. to meet Trump. A bipartisan group of senators pressed Apple to abandon Chinese memory chip purchases, warning the company would increase dependence on an adversary nation for critical components.

I.Top Stories

Altman works the White House and the Senate before the Aug. 1 frontier AI framework deadline

OpenAI CEO Sam Altman is meeting senior Trump administration officials, lawmakers and economists this week, including White House chief of staff Susie Wiles, days before the Aug. 1 deadline for the administration's frontier AI model review framework, CNBC reported. Sen. Ted Cruz, R-Texas, and several Senate Democrats met with Altman on Wednesday. In a separate round of Washington meetings in early June, Altman met Michael Kratsios, the White House's chief science and technology adviser, congressional leaders from both parties.

Read at CNBC ↗ • Read at AP ↗

METR, Redwood launch independent review of OpenAI Hugging Face incident

METR said it reached an agreement with OpenAI to conduct an independent review, with Redwood Research, of the model behavior observed during the Hugging Face incident. The review will be limited to a specific set of questions, and METR plans to publish the terms of the engagement, the scope covered and tentative conclusions. OpenAI reported last week that some of its internal frontier agents autonomously hacked into Hugging Face in an attempt to access the answer key for a cybersecurity benchmark. In a July 28 post, METR said a fuller propensity investigation would require the ability to run the models involved, access to full transcripts, interviews with security, training and internal investigation staff, and the ability to run prompted classifiers over training data. METR said conclusions should go to the company's board and be made public subject to redactions, with a redaction summary describing what the process left unsubstantiated.

Read at X ↗ • Read at METR ↗

HAWK developer withdraws post-quantum signature scheme after AI-assisted attack halves key strength

The developer of HAWK, a digital signature scheme that had cleared two rounds of NIST review, withdrew it Tuesday, a day after Anthropic published results from its Mythos model, Ars Technica reported. HAWK's security rests on the hardness of the Lattice Isomorphism Problem, which, unlike the digital signature schemes in widest use today, is believed to resist attacks from quantum computers; the scheme was in a third round of testing designed to surface exactly this kind of flaw. An Anthropic researcher with no cryptography expertise spent about 60 hours and roughly $100,000 in computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. prompting Mythos to improve the best known existing attack, cutting HAWK's effective key strength in half. The weakness can be mitigated by doubling the key size, but the added computation leaves HAWK less competitive than ML-DSA and FN-DSA, two post-quantum signature schemes already available. Sophie Schmieg, a post-quantum cryptography specialist at Google, wrote that "basically with this paper, HAWK is dead."

Read at Ars Technica ↗

Beijing threatens retaliation over the FCC ban on Chinese humanoid robots and power inverters

China's Commerce Ministry said Thursday it will retaliate against the Federal Communications Commission's ban on imports of new foreign-made humanoid robots, robot dogs and power inverters, CNBC reported. The ministry said the measure "severely damages" relations and that the FCC has repeatedly ignored Beijing's restrained stance. Chinese manufacturers account for roughly 85% of the global humanoid robot market, and Morgan Stanley analysts forecast China's market for humanoids could reach $15 billion by 2030, according to the AP. Samm Sacks, a senior fellow at the New America think tank focused on Chinese technology policies, described the exchange as a steady series of flashpoints building toward the September summit. Xi Jinping is expected to visit the U.S. that month to meet Trump.

Read at CNBC ↗ • Read at AP ↗

Senators tell Apple to drop its plan to buy memory chips from CXMT and Yangtze Memory

A bipartisan group of senators urged Apple to abandon any effort to buy memory chips from CXMT and Yangtze Memory Technologies, warning that the company would become reliant on a U.S. adversary for a critical component, Bloomberg reported. Both suppliers sit on the Pentagon's list of Chinese military companies, and Yangtze Memory has been on the Commerce Department's Entity List, which restricts U.S. technology exports to named firms, since 2022. Apple has been negotiating with the two for memory destined for devices sold in China, a deal that would take its memory roster to five suppliers alongside Samsung, SK Hynix and Micron, The Straits Times reported. Memory prices have climbed as AI data center construction absorbs supply, and Apple has raised prices across its Mac, iPad, home device and Vision Pro lines.

Read at Bloomberg ↗ • Read at The Straits Times ↗

OpenAI says the research agent behind the Hugging Face breach also reached four other services

OpenAI updated its incident post on Tuesday to say the pre-release research agent behind the Hugging Face compromise used credentials it found online to reach four accounts on four other publicly available services, per The Verge. The company said it has "not identified any other activity at the level of severity or scale" of the Hugging Face incident, which it described as a platform-level compromise. Modal Labs was among the services reached, though Modal's own infrastructure was not compromised. OpenAI said the system was an internal-only research prototype that has been deactivated, encrypted and restricted, and that it will publish a technical report in the coming weeks.

Read at The Verge ↗ • Read at OpenAI ↗

Safety researchers read the OpenAI agent episode as an argument for pre-release model review

Researchers assessing OpenAI's disclosure about the Hugging Face compromise said the episode strengthens the case for reviewing frontier models before release, The Verge reported. Fazl Barez, an AI safety researcher at the University of Oxford, placed the behavior in a category researchers call specification gaming, which he described as "the model doing what you asked rather than what you meant." Seán Ó hÉigeartaigh, a professor at Cambridge University's Leverhulme Centre for the Future of Intelligence, called the incident "a pretty useful warning shot."

Read at The Verge ↗

II.China Watch

China's internet regulator drafts a cyberbullying law that reaches AI generated abuse

The Cyberspace Administration of China, the agency that polices online content there, released a draft Anti-Cyberbullying Law for public comment on Wednesday, per Caixin. The 60-article text, written with other government departments, defines cyberbullying as concentrated or sustained online attacks on someone's reputation, privacy, likeness or personal information. It requires platforms to detect and trace abuse produced with AI tools. Most of the draft covers platform duties: it bars bulk or fake account registration, trading in accounts, manufacturing false trending topics and manipulating rankings and hot search lists, and it sets additional obligations for the largest platforms. It also establishes court injunctions to halt ongoing harassment, a route for turning a private criminal complaint into a public prosecution, and a requirement that providers alert users to emerging pileups and report cases involving minors, older adults or people with disabilities. The Standing Committee of the National People's Congress, which passes China's laws, listed cyberbullying legislation as a preparatory item in its 2026 legislative plan on May 11. Comments on the draft are due Aug. 28.

Read at Caixin ↗

ByteDance folds its workplace software team into its chatbot and cloud units to chase enterprise AI customers

ByteDance restructured its AI business on Thursday, splitting the team behind Feishu, its workplace collaboration suite, between the group that builds its Doubao chatbot and Volcano Engine, its cloud and model service arm, per TechNode. Feishu's product staff move into a new Doubao product organization, while its sales, marketing and support staff join Volcano Engine's in an enterprise facing unit called the Creativity Service Platform. ByteDance said Feishu's existing products and services will not change and that the reshuffle is meant to tighten coordination on workplace AI. Caixin reported that Feishu head Xie Xin now reports to Doubao head Zhao Qi. ByteDance put the annualized revenue run rate of its model business at $4 billion based on average consumption in July, a figure the company said exceeds the combined total of every other Chinese model developer. A Doubao enterprise edition built largely by the Feishu product team has already entered internal testing with selected Feishu customers.

Read at TechNode ↗

A Tsinghua spinout says it can knit distant Chinese data centers into a single AI service

Infinigence AI, a startup founded out of Tsinghua University that sells software for running models across mismatched chips, published the full technical report on Thursday for a system that splits one model's work across data centers in different cities, per QbitAI. The system begins generating an answer over a fast local link while a distant cluster is still receiving the context over ordinary internet lines, then hands the rest of the job to that cluster. Infinigence said the aim is to put already built clusters in different provinces to work together, and to let mismatched chips handle only the stage they are fast at rather than buying costly mixed hardware for one site. In the company's own measurements the wait for the first word of a response fell by about half, and the cost of generating each token, the unit of text that model use is billed by, fell 37.5%. QbitAI was the only outlet reporting the results, and no independent testing has been published. The code and the report are posted on GitHub, and the architecture was first shown at the World Artificial Intelligence Conference in Shanghai, which closed this month.

Read at QbitAI ↗

A Chinese security firm puts an autonomous bug hunting agent into its products, running on a domestic open model

Sangfor, a Chinese network security company, said Wednesday that its Sangfor AI agentAI 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. completed 1,301 of 1,507 tasks on CyberGym, a public benchmark that asks systems to rediscover and reproduce known security flaws in open source software, per QbitAI. The agent ran on Zhipu AI's GLM-5.2, an open weight Chinese model, under the benchmark's fixed model setting. Sangfor said the score placed it fourth worldwide and first among Chinese entrants, ahead of entries from OpenAI and Anthropic; CyberGym publishes its leaderboard publicly. The agent runs several lines of investigation in parallel and submits a candidate exploit only after checking that the crash it found can be reproduced and matches the flaw being targeted. Sangfor said the system is already running inside its enterprise code security products, extending them from rule matching to logic flaws such as broken access control and authentication bypass.

Read at QbitAI ↗

III.Policy Tracker

House Energy and Commerce's ranking member asks SpaceX for records on its Memphis AI data centers

Rep. Frank Pallone, D-N.J., ranking member of the House Energy and Commerce Committee, asked SpaceX on Wednesday for records on its Memphis area AI data centers and for a visit to the sites and the power plants serving them, CNBC reported. The sites are Colossus 1 and Colossus 2, the computing clusters SpaceX runs for xAI, which SpaceX acquired earlier this year. Residents living near them have complained about noise and emissions from the natural gas turbines supplying their power.

Read at CNBC ↗

Two senators introduce the RESULTS Act to rewire state education and workforce data for AI era skills

Sens. Lisa Blunt Rochester, D-Del., and Roger Marshall, R-Kan., introduced the RESULTS Act on Wednesday, StateScoop reported. It would rewrite the federal Statewide Longitudinal Data Systems program, which funds state databases linking school records to later employment. The bill would broaden grant eligibility to governors and state data governance bodies, open access to the National Directory of New Hires and merge the Education Department's program with the Labor Department's Workforce Data Quality Initiative. Blunt Rochester said AI presents new opportunities for education and the economy but that decision makers need better access to reliable data. Marshall said AI is a tool rather than a replacement for the American worker. Thirty-three states run such systems and nine are building them. The Education Department's statistics arm has awarded more than $930 million in program grants since 2005.

Read at StateScoop ↗

Senate Aging Committee takes testimony on AI voice clones and deepfakes in elder fraud

The Senate Special Committee on Aging held a hearing Wednesday titled "The AI Deception Machine: DeepfakesDeepfakeSynthetic 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., Chatbots, and the New Frontier of Senior Fraud." Chairman Rick Scott said Americans over 60 lost $7.7 billion to scams in 2025. Ranking Member Kirsten Gillibrand noted she has written to leading AI companies, including OpenAI and Meta, seeking details on safeguards for older adults and people with disabilities. Witness Deborah Del Mastro testified about a five-and-a-half-hour ordeal in which scammers used a clone of her daughter's voice to stage a kidnapping. Senators discussed bipartisan legislation for a national anti-scam strategy.

Read at NewKerala ↗ • Read at aging.senate.gov ↗

Commerce Department signs a $300 million letter of intent with GlobalFoundries and takes a 1% stake

The Commerce Department signed a letter of intent Wednesday to provide GlobalFoundries $300 million for silicon photonics research, with the federal government taking a 1% equity stake in the chipmaker in exchange, SiliconANGLE reported. The money targets co-packaged optics, which places optical components in the same package as a processor so data moves between AI accelerators over light rather than copper. GlobalFoundries introduced its SCALE co-packaged optics technology in May. The $300 million follows a separate $375 million federal grant to the company two months ago for a quantum chip foundry.

Read at SiliconANGLE ↗

DHS plans more AI redaction in FOIA processing as its pending caseload grows

The Department of Homeland Security said in its annual Freedom of Information Act report, published Thursday, that it expects to launch more automation tools in the coming year to process records requests, FedScoop reported. The agency's inventory of its AI systems describes a tool that would speed the identification and redaction of sensitive content. The project is still in pre-deployment and classified as not high impact, which exempts it from the stricter risk management rules applied to high impact government AI. DHS opened fiscal 2025 with 221,068 pending requests and ended it with 245,572 after receiving more than 1 million. Abigail Kunkler, a law fellow at the Electronic Privacy Information Center, said automating redactions "almost guarantees that DHS will over-redact." DHS last month posted a notice of intent to award Deloitte a contract for FOIA software with AI tools.

Read at FedScoop ↗

IV.Capability & Research Watch

Google's SynthID watermark holds up in testing, but stripping AI labels stays easy

Ars Technica tested SynthID, Google's system for embedding an invisible watermark in AI generated images, and found the detector accurate on unaltered files but the watermark easier to defeat once an image is edited or re-encoded. Google said in the spring that its tools had been used to create more than 100 billion AI images and videos. OpenAI, Runway and Nvidia have begun adopting SynthID. Pushmeet Kohli, a Google DeepMind scientist, said the team assumed from the start of development that the technology would be attacked. The competing standard, C2PA provenance metadata, is cryptographically signed but disappears when an image is screenshotted or re-exported.

Read at Ars Technica ↗ • Read at Nature ↗

FAR.AI report finds 448 working jailbreaks against Grok and 249 against Gemini

FAR.AI, a California AI safety nonprofit, generated more than a thousand variants of harmful prompts against seven frontier models from four companies, Wired reported. It found 448 working jailbreaks against SpaceXAI's Grok and 249 against Google's Gemini 3.1 Pro, while OpenAI's GPT models and Anthropic's Claude and Fable resisted the attacks. The computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. bill came to $58 for Grok and $278 for Gemini. Adam Gleave, FAR.AI's chief executive, said "AI models right now are less regulated than restaurants" and said voluntary industry commitments are not a workable substitute for standards. Rohin Shah, Google DeepMind's director of AGIAGIArtificial general intelligence: an AI system that can do most economically valuable cognitive work at or above human level. There is no agreed test for it, which is why debates about when it will arrive, and what to do about it, are so contentious. safety and alignmentAlignmentThe problem of making an AI system reliably pursue the goals its developers and users intend, and the research field devoted to it. Misalignment covers everything from a chatbot flattering users to a capable system deceiving or resisting its operators., said the results should not be read as a full assessment of Gemini's safety and security. California and New York require frontier developers to publish safety reports, and an Illinois law will require third party audits of safety practices. No federal safety requirement exists.

Read at FAR.AI ↗ • Read at WIRED ↗

V.Industry & Market Watch

Brookfield and NextEra plan a $100 billion AI data center on the Energy Department's Paducah site

Brookfield Asset Management and NextEra Energy said Wednesday they will develop a $100 billion AI data center campus in western Kentucky on the Energy Department's Paducah site, a Cold War-era uranium enrichment plant, The Wall Street Journal reported. The partnership includes the nonprofit utilities Big Rivers Electric Power and Jackson Purchase Energy Cooperative along with the municipal Paducah Power System. The developers said the project is privately funded.

Read at The Wall Street Journal ↗

Microsoft signs $130 billion in data center leases as intelligent cloud revenue rises 32%

Microsoft has signed $130 billion of data center leases, and revenue in its intelligent cloud segment rose 32% to $39.3 billion in the quarter, the Financial Times reported. Total revenue was $90.01 billion, up 18%, and net income was $35.77 billion. Earnings of $4.74 a share included about 27 cents from accounting items, mostly a $3.2 billion unrealized gain on Microsoft's stake in Anthropic after that company's valuation rose from $350 billion to $900 billion, per SiliconANGLE. Capital spending and finance leases reached $41 billion in the quarter, up 69% from a year earlier. Microsoft also plans to extend the assumed useful life of its new offices and data center buildings from 15 to 25 years, spreading depreciation over a longer period.

Read at Financial Times ↗ • Read at SiliconANGLE ↗

Meta's AI buildout leaves $784 million of free cash flow as profit falls 14%

Meta reported revenue of $60.80 billion for the quarter ended June 30, up 28%, SiliconANGLE reported. Net income fell 14% to $15.85 billion as total costs rose 55% to $42.03 billion. Research and development spending rose 67% to $21.66 billion, and capital spending on the AI buildout, including finance lease principal, reached $31.08 billion. Free cash flow was $784 million, against $8.55 billion a year earlier. Meta ran no share buybacks in the quarter, compared with $10.17 billion a year earlier, and long-term debt rose to $83.66 billion from $58.74 billion. Shares fell more than 8% in after-hours trading.

Read at SiliconANGLE ↗

Visa cuts about 2,600 jobs, 7% of its workforce, in a pivot toward AI

Visa is cutting about 2,600 jobs, roughly 7% of its workforce, mostly in technology and product teams, Silicon Republic reported. Chief Executive Ryan McInerney told staff that "AI is also helping to accelerate this evolution and shape the way work gets done at Visa." People familiar with the plan told Bloomberg that AI efficiency was not the only driver, with money moving toward consumer payments, stablecoins, cross-border transfers and business-to-business services. Visa's net revenue rose 14% to $11.6 billion last quarter. Layoffs.fyi counts 252 technology companies cutting more than 124,000 jobs so far in 2026, above the roughly 123,000 recorded in all of 2025.

Read at Silicon Republic ↗

VI.Global & Geopolitics

ChatGPT and Roblox cross the EU threshold that triggers its strictest platform oversight

OpenAI's ChatGPT and Roblox are set to fall under the European Union's strictest platform obligations after each passed 45 million average monthly users in the bloc, Bloomberg reported. That threshold triggers designation as a very large online platform under the Digital Services Act, the EU's content moderation law, which is separate from the AI Act. The announcement is expected as soon as August. Designated services must assess systemic risks such as illegal content and harm to minors every year, submit to independent audits and answer to the European Commission directly rather than to a single national regulator. They have four months from designation to comply.

Read at Silicon Republic ↗ • Read at Bloomberg ↗

India's IT ministry says generative AI services qualify for safe harbor case by case

Jitin Prasada, India's minister of state for electronics and information technology, told the Lok Sabha, the lower house of India's parliament, in a written reply Wednesday that whether a generative AI service counts as an intermediary eligible for the liability shield in section 79 of the Information Technology Act "depends upon the nature of the service provided, the functions performed," rather than on a blanket rule. The ministry said the act is technology neutral and applies to AI services on the same terms as other online services. Amendment rules under the act, issued Feb. 10 and in force since Feb. 20, require platforms to label synthetically generated content, including deepfakesDeepfakeSynthetic 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., and to embed traceable metadata. Providers that fail to comply lose the section 79 protection.

Read at NewKerala ↗