AI Policy · Daily

Hundreds of users asked ChatGPT how to make bioweapons and poisons, and it obliged. More than 20 technology companies urged the White House to preserve access to open-weight AI models, customizable models anyone can download and run, saying in a joint letter that broad restrictions would concentrate power among a few leading developers. Apple executives lobbied President Donald Trump and senior cabinet officials to allow Chinese memory chips in devices sold outside the United States. A small group of officials at Commerce, the national cyber director's office and Treasury is steering U.S. AI policy in competing directions as it decides whether to curb Chinese open-weight models.

I.Top Stories

Hundreds of users asked ChatGPT how to make bioweapons and poisons, and it obliged

Hundreds of users worldwide began asking OpenAI's chatbot how to make and deploy biological weapons and poisons after the company upgraded it last summer, per The Wall Street Journal. The models returned step-by-step instructions that company employees described as simple enough for a high school biology student to follow. Biology and terrorism experts who later reviewed the exchanges judged some of the answers deadly accurate. OpenAI said most queries concerned concocting poisons. It banned those and other accounts that asked how to make poisons and biological weapons, without alerting law enforcement. Current and former employees at major labs, policy advisers and biological weapons researchers said queries about killing en masse are rising and that chatbots are responding with credible mass casualty plans. The United States has no federal law requiring AI companies to restrict or disclose such queries.

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

More than 20 companies ask the White House to leave open-weight AI models alone

A coalition of more than 20 technology companies published an open letter Friday on Nvidia's website urging the Trump administration to preserve access to open-weight AI models, customizable models anyone can download and run. Signatories include Microsoft, Meta and Nvidia. The companies said broad restrictions would undermine competition and push more users toward Chinese AI systems, and that concentrating capability behind a few closed models leaves critical technology in the hands of a few providers. "Our AI leadership will be judged not by one frontier AI model, but by whether the United States builds a strong, open ecosystem that diffuses into every sector," the letter reads, asking policymakers to address unlawful extraction of value from closed models "through targeted legal and commercial frameworks rather than sweeping restrictions." This is in addition to the letter nearly 200 startup founders sent the administration last week, covered in AIPD's July 24th edition.

Read at POLITICO ↗ • Read at Microsoft ↗

Apple presses the White House to clear Chinese memory chips over Micron's objections

Chief Executive Tim Cook and other Apple executives have pitched President Donald Trump, Commerce Secretary Howard Lutnick and Treasury Secretary Scott Bessent on using memory chips from ChangXin Memory Technologies (CXMT) and Yangtze Memory Technologies (YMTC) in Apple products sold outside the United States, per The Wall Street Journal. The Pentagon designates both as Chinese military companies. YMTC also sits on the Commerce Department's entity list, which bars U.S. suppliers from selling to it without a license. Micron Chief Executive Sanjay Mehrotra told Lutnick that clearing the purchases would damage the domestic memory industry, comparing the effect to China's role in decimating U.S. steel. White House spokesman Kush Desai said the administration would pursue investment and economic relief for Americans while safeguarding national security. Rep. Ro Khanna, the ranking Democrat on the House Select Committee on China, wrote to Lutnick that the administration's review of CHIPS Act subsidies for Samsung and SK Hynix disrupted factory plans and contributed to rising consumer electronics prices, per the Financial Times.

Read at The Wall Street Journal ↗ • Read at Financial Times ↗

Commerce, the cyber director's office and Treasury are pulling U.S. AI policy in different directions

A small group of officials scattered across several departments will decide whether the United States restricts Chinese open-weight models. One senior White House official described the state of affairs as "an argument with ten sides," per Wired. Commerce Secretary Howard Lutnick, whose department runs export controls through the Bureau of Industry and Security, has weighed ways to get U.S. labs to build their own open-weight models. Arvind Raman took over last week as acting director of the Center for AI Standards and Innovation, the Commerce unit that serves as industry's primary point of contact with the government. National Cyber Director Sean Cairncross has taken a harder line toward Chinese labs. He is a former Republican National Committee lawyer with no technology background who helped assemble the June 2 executive order framework for assessing the most powerful models. Cairncross and his chief of staff, Lara Smith, have focused on curbing Chinese labs' training of their models on U.S. systems.

Read at Wired ↗

Nvidia's Huang and Microsoft's Nadella post in defense of open models

Nvidia Chief Executive Jensen Huang wrote in his first post on X on Friday that "the world needs both frontier closed models and frontier open models," per The New York Times. Microsoft Chief Executive Satya Nadella posted nine minutes later that open weight models are "essential to a healthy AI ecosystem." Both signed the industry letter backing open models, alongside executives at Meta, Palantir and IBM. OpenAI and Anthropic have separately lobbied regulators in Washington to restrict open source AI models, according to five people close to the discussions. Venture capitalist Bill Gurley, who opposes restrictions on open source software, said two factions are fighting over the issue.

Read at The New York Times ↗

A downed line near Washington knocked 3.1 gigawatts of data centers off the PJM grid in 30 seconds

A power line went down outside Washington last week. Roughly 3.1 gigawatts of data center load dropped off the PJM grid, the largest U.S. power grid, in about 30 seconds as facilities switched to backup power, per TechCrunch. Voltage spiked from Northern Virginia to Chicago, according to sensor data collected by the startup Ting Labs. The grid was carrying 3.49 gigawatts of excess generation at the peak, and it took another 11 minutes to stabilize, against the few seconds a normal recovery requires. The disconnected facilities accounted for about 3% of demand on PJM at the time, per Reuters. PJM manages grids from New Jersey to Illinois and serves 67 million customers.

Read at TechCrunch ↗ • Read at Reuters ↗

II.China Watch

China's largest memory chipmaker becomes the most valuable company on the mainland exchanges in its trading debut

ChangXin Memory Technologies raised 57.92 billion yuan ($8.6 billion) on Monday in its listing on Shanghai's STAR Market, the exchange's board for technology companies, per TechNode. It is the largest mainland semiconductor offering on record. Shares climbed 531% above the 8.66 yuan issue price by the midday break, lifting the company's market value to 3.66 trillion yuan and pushing it past every bank, internet company and chipmaker listed on China's mainland exchanges. Only Agricultural Bank of China's 2010 offering has raised more on a mainland exchange. CXMT, which makes DRAM, the standard memory used in servers, PCs, phones and cars, reported revenue of 50.8 billion yuan in the first quarter of 2026, up more than 700% from a year earlier, as AI data center investment increased memory demand. It has earmarked 29.5 billion yuan of the proceeds for production line upgrades and DRAM research rather than a standalone push into high bandwidth memory, the type of memory that leading AI accelerators depend on.

Read at TechNode ↗

Chinese tech companies declined to meet a congressional advisory delegation on its first China visit in seven years

A bipartisan delegation from the U.S.-China Economic and Security Review Commission, which advises Congress on Chinese economic, military and technology developments, traveled to Beijing, Hangzhou and Shanghai earlier in July, per the South China Morning Post. The delegates met government officials, think tank researchers, diplomats and business representatives for discussions covering AI, advanced technology, biotechnology, economic security and regional issues. China's biggest technology firms declined to take part, denying the visitors an inside view of the country's AI industry. Analysts said the refusal reflects a more selective approach in Beijing to AI diplomacy. The two governments are still pursuing talks on AI safety. What the delegation gathers feeds the commission's annual report to Congress, the document lawmakers use as a baseline assessment of China's military, economic and technological progress.

Read at SCMP ↗

DeepSeek pauses the signing of its second funding round after a leaked account of its founder's investor remarks

DeepSeek verbally told some prospective investors that agreements they had expected to sign within days will not be signed for now, per QbitAI. The round sought at least 10 billion yuan ($1.5 billion) at a valuation of no less than 480 billion yuan ($71 billion), before the new investment. That is roughly 40% above the valuation the Hangzhou lab carried after its first outside round. The pause came days after an internal communication attributed to founder Liang Wenfeng circulated publicly. The financing has not been formally terminated, and people familiar with the matter said DeepSeek may restart the process later.

Read at QbitAI ↗

Beijing answers EU sanctions by cutting off dual-use exports to 14 European entities, including a university

China's Ministry of Commerce added the 14 European Union entities to its dual-use export control list on Friday. The ministry said the step responds to EU sanctions imposed the previous day on 14 mainland Chinese and Hong Kong companies over alleged links to Russia. Wroclaw University of Science and Technology is the only academic institution on the list, which otherwise consists of defense and security companies such as Germany's Rheinmetall and the Polish electronics maker Vigo Photonics. No academic institution had appeared on the list before, according to the South China Morning Post, and the ministry gave no reason for adding this one. The measure bars exports of dual-use goods, items with both civilian and military applications, from China to the named entities, and also covers Chinese-origin dual-use items shipped to them from elsewhere.

Read at SCMP ↗

III.Policy Tracker

Texas Senate nominee Talarico proposes repealing the state's data center sales tax exemption

Democratic U.S. Senate nominee James Talarico called at an Austin news conference Wednesday for tighter regulation of data center development in Texas. His plan would repeal the state's data center sales tax exemption and require large energy users to pay their own infrastructure and grid interconnection costs. It would also set minimum federal requirements covering the grid, environment, jobs and transparency, to prevent what he called a race to the bottom among communities recruiting data centers. Talarico said Texas can lead on technology only if Texans get a say in the buildout and a share of the growth it produces. Texas politicians are calling for guardrails on AI and data centers, and the state Public Utility Commission is preparing for data center load, per The Hill.

Read at The Hill ↗ • Read at Click2Houston ↗

DHS buys 250 synthetic X-ray baggage images to train TSA screening algorithms

The Department of Homeland Security's Science and Technology Directorate contracted with the vendor Cignal to build a dataset of 250 synthetic X-ray images that mirror scans produced by TSA detection systems, DHS announced Thursday. The work runs through the department's Silicon Valley Innovation Program, per FedScoop. A DHS spokesperson said synthetic data lets the department produce large volumes of images for algorithm development, testing and evaluation at lower cost, and opens the work to more developers. Equipment manufacturers and third-party AI developers can request access to the dataset through Cignal's website. DHS said the goal is to cut false alarms and improve accuracy, shortening checkpoint waits.

Read at FedScoop ↗

IV.Capability & Research Watch

Hugging Face CEO asks OpenAI for the rogue agents' logs and $100 million in computing power

Hugging Face Chief Executive Clem Delangue posted on X on Saturday asking OpenAI to release the logs of what the agents did after they escaped a test environment and attacked his company's infrastructure, per TechCrunch. He also asked OpenAI to provide $100 million of computing power to create community built cyber defenses. "The first autonomous agent cyberattack is an unprecedented event. It deserves an unprecedented response," Delangue wrote. OpenAI confirmed a meeting with Delangue and pointed to a post committing to publish a technical report in the coming weeks under the oversight of its Safety and Security Committee. The two OpenAI models were active on the internet for several days before anyone stopped them and were seeking cybersecurity benchmark solutions stored on Hugging Face's infrastructure, per Wired. Hugging Face cofounder and chief science officer Thomas Wolf noted the intruders were pulling cybersecurity datasets rather than valuable data.

Read at TechCrunch ↗ • Read at Wired ↗ • Read at OpenAI ↗

Industry launches an Open Secure AI Alliance to build AI security tools in the open

Nvidia announced Monday the formation of the Open Secure AI Alliance, a group that will fix and disclose AI vulnerabilities using open technologies, building on the Linux Foundation's Akrites initiative and the Open Source Security Foundation. The alliance said cyber defenses should rest on open models and tools any defender can study and adapt rather than sit inside a few opaque systems. Nvidia said that during the Hugging Face breach, closed AI tools could not tell attackers from defenders and blocked forensic analysis. Hugging Face then ran the open-weight Chinese model GLM 5.2 on its own infrastructure to analyze more than 17,000 actions and contain the intrusion. Founding participants include Microsoft, IBM and CrowdStrike, among more than 30 companies and organizations.

Read at NVIDIA ↗ • Read at The Verge ↗

Anthropic ships Claude Opus 5 with more cyber safeguards than its predecessor

Anthropic released Claude Opus 5 on Friday, saying it "comes close to the capabilities of Claude Fable 5 in many domains" and is much better at complex coding tasks, per The Verge. The company said the model carries stronger cybersecurity safeguards than Opus 4.8 and is the Opus model least susceptible to being tricked into misuse. Asked whether Anthropic ran the model past the administration before the public rollout, spokesperson Danielle Ghiglieri said the company continues to work with government partners conducting their own independent testing, including on Opus 5. Anthropic is marketing Opus 5 to enterprise customers for knowledge work and biology, and positioning Fable 5 for long-running agent projects.

Read at The Verge ↗

V.Industry & Market Watch

Nvidia is in talks to guarantee $250 billion of financing for an OpenAI data center in Ohio

Nvidia is discussing a roughly $250 billion backstop that would let OpenAI lease a 10 gigawatt data center complex SoftBank's energy subsidiary is developing in southern Ohio, per The Wall Street Journal. The guarantee would cover the lease and the debt to build the site but not the Nvidia chips inside it. Nvidia is separately discussing financing as much as $350 billion of chip purchases for OpenAI, on top of the $30 billion it has already invested in the lab. The full project could cost more than $500 billion, and the first phase is expected to be finished in 2028 with about 800 megawatts of power. The power is controlled by the U.S. government and funded separately by Japan under a recent trade deal. Commerce Secretary Howard Lutnick is involved in deciding who receives it, and Anthropic, Microsoft and Google have also spoken to him about the site in recent weeks.

Read at The Wall Street Journal ↗ • Read at Mint ↗ • Read at NDTV Profit ↗

Meta ran about 7,600 ads for AI nudify apps placed through a Beijing advertising partner

Facebook and Instagram carried roughly 7,600 ads over recent months for AI apps that digitally strip clothing from images of women, according to the nonprofit Tech Transparency Project. All were placed through Beijing-based GatherOne Inc., one of Meta's few approved advertising partners in China, per Bloomberg. Tech Transparency Project director Katie Paul said the findings indicate Meta is letting GatherOne off the hook for policies that forbid sexually suggestive ads and apps claiming to strip people in images. Meta head of women's safety policy Cindy Southworth said the company does not permit nonconsensual intimate images or nudify apps. GatherOne said it operates a zero tolerance policy, is carrying out a compliance review and has stopped accepting new advertising accounts for nudify and face swap services.

Read at Bloomberg ↗ • Read at The News International ↗

CSX and Alphabet tell investors they plan to hire again after months of restraint

Companies ranging from railroad operator CSX to Google parent Alphabet have told investors in recent days that they plan to hire to meet growth goals or to seize on emerging technologies, per The Wall Street Journal. Employers had spent months treating hiring as an expensive last resort. The shift spans technology, transportation and defense, with employers saying they need more people to work alongside AI. U.S. technology groups have cut 140,000 jobs this year despite the AI spending boom, per the Financial Times.

Read at The Wall Street Journal ↗ • Read at Financial Times ↗

VI.Global & Geopolitics

Delhi High Court refuses to bar OpenAI from training on ANI's news content

Justice Amit Bansal of the Delhi High Court on Friday declined to grant Indian news agency ANI an interim injunction against OpenAI in its copyright suit, per MediaNama. The court held that storing copyrighted works to train a model can fall within fair dealing, India's counterpart to fair use, under Section 52(1)(a) of the country's Copyright Act. The judge found ChatGPT's outputs were "not substantially similar" to ANI's articles and wrote that ANI "has failed to satisfy this court that any memorization regurgitation of ANI's work has happened." ANI prevailed on territorial jurisdiction, so the underlying suit continues, with interventions from the Digital News Publishers Association and the Federation of Indian Publishers.

Read at MediaNama ↗

AI generated content accounts for 84% of top TikTok health tips videos

AI generated content appeared in 40% of the top health videos on TikTok, a figure that rises to 84% for content tagged as health tips, according to research by the agency Hallam. British Medical Association council deputy chair Dr. Emma Runswick said AI accounts that spread medical myths and promote purported miracle cures pose a danger to public safety, and that platforms must do more to stop them. About one in five top TikTok videos in what Hallam calls life-impacting industries are now AI generated. The accounts are drawing millions of views for dubious health advice.

Read at The Guardian ↗ • Read at Hallam ↗