AI Policy · Daily

Security researchers at Hacktron AI used Anthropic's Claude to reach OpenAI's internal code, entering an employee's ChatGPT account in a bug bounty test that drew a $6,500 payout. The House left Washington to campaign without voting on AI legislation, with no session days left until after the Nov. 3 midterms and the heads of Anthropic, OpenAI and xAI pressing for federal oversight the same week. The New York Times editorial board urged Congress to license AI developers, proposing an Artificial Intelligence Commission modeled on the Atomic Energy Commission and funded through industry fees. Pew Research Center found a median of 43% across 12 countries trust China to regulate AI effectively, against 35% for the U.S. and 34% for the EU.

I.Top Stories

Security researchers use Anthropic's Claude to reach OpenAI's internal code

Independent security researchers used Anthropic's Claude to get into an OpenAI employee's ChatGPT account, giving them a way to read and suggest changes to the company's private store of code, per the Wall Street Journal, which describes the group as bug hunters rather than attackers. The researchers work at the security startup Hacktron AI and reached OpenAI's internal code on GitHub through the employee's account, per Forbes. OpenAI paid them a $6,500 bounty after they reported the flaws through its bug bounty program, which rewards outsiders for disclosing security weaknesses. Hacktron put the cost of the exercise at less than $3,000 in tokens.

Read at Wall Street Journal ↗ • Read at Forbes ↗

House adjourns through Nov. 3 elections without taking up AI legislation

The House left Washington to campaign with no AI legislation passed and no voting days left until after the elections, per CNBC. The heads of Anthropic, OpenAI and xAI had pressed Washington for federal oversight during the same week, and Trump dismissed the campaign to limit the technology as a conspiracy, per AP. Speaker Mike Johnson has left the issue to the White House, while Sen. Bernie Sanders (I-Vt.) convened a briefing with outside experts for senators on Wednesday.

Read at AP ↗ • Read at CNBC ↗

New York Times editorial board urges federal AI commission with licensing power

The New York Times editorial board called on Congress to create an Artificial Intelligence Commission modeled on the Atomic Energy Commission, which President Harry Truman signed into law on Aug. 1, 1946, and to require AI developers to obtain a federal license. Licenses would turn on three principles: a government-written "constitution" setting the values each model must respect, disclosure and watermarking of AI interactions and content, and mandatory independent safety testing before release. The board said the voluntary slowdown raised by Dario Amodei, Sam Altman and Elon Musk is welcome but insufficient, partly because coordination among companies raises antitrust issues. It also proposed funding the agency through industry fees and, citing Yoshua Bengio, creating a separate accident investigator modeled on the National Transportation Safety Board.

Read at The New York Times ↗

Pew poll finds more trust in China than U.S. or EU to regulate AI across 12 countries

A median of 43% of respondents said they trust China to regulate AI effectively, against 35% for the U.S. and 34% for the EU. The Pew Research Center put the question to 13,696 adults across 12 mostly middle income countries and territories between Feb. 8 and May 13. China ranked most trusted in five of them, including Bangladesh, Malaysia and Pakistan, while the Philippines was the only place where the U.S. placed first. Respondents in Chile and Thailand trusted the three about equally. China was not itself among the countries surveyed on the question.

Read at Pew Research Center ↗

Palantir's Karp floats nationalizing top AI companies, wants executives liable first

Palantir chief executive Alex Karp told CNBC that the risks from AI are high enough that the leading companies may have to be brought under government ownership. He said the industry should have enforceable guidelines, but that the first recourse should be holding companies and their executives answerable for what their systems do, per The Motley Fool. Tech builders who are not being responsible should face civil and criminal liability, Karp said. He said the push for government oversight is partly an attempt to shield the companies from that liability.

Read at CNBC ↗ • Read at The Motley Fool ↗

II.China Watch

U.S. Entity List pushed sanctioned Chinese firms deeper into published research, Science study finds

Chinese companies added to the U.S. Entity ListEntity ListA Commerce Department blacklist of foreign companies and organizations that US firms may not supply with controlled goods without a license. Adding Chinese chipmakers and AI labs to the list is the most common enforcement step in the technology export fight., which bars them from receiving controlled American technology without a license, later produced 72.3% more patents citing at least one scientific publication than comparable unlisted firms, the study found. Junhan Wang, Xibao Li and Yanbo Wang published the analysis in the journal Science. Listed firms put out more papers of their own at home and abroad, 33.3% more indexed in the China National Knowledge Infrastructure and 85.2% more in Web of Science. The authors wrote that U.S. trade sanctions such as the Entity List have generated consequences U.S. policymakers may not have fully anticipated. The study covers Chinese patent filings and corporate publications from 2010 through 2022.

Read at South China Morning Post ↗ • Read at Nature ↗

Compute to take 6% of China's electricity by 2030, power industry council projects

National electricity use by internet data services has passed 60 billion kilowatt-hours so far this year, up more than 40% from a year earlier, the China Electricity Council said in a Sept. 16 release, per Caixin. Council supervisor Pan Yuelong said electricity consumed by computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. Because it is expensive, physical and concentrated in a few suppliers, compute is the main lever governments pull to shape who can build frontier AI. will rise by more than 100 billion kilowatt-hours a year on average during the 15th Five-Year Plan, China's national economic blueprint for 2026 through 2030. That would reach about 800 billion kilowatt-hours by 2030, roughly 6% of all power used in the country. Fujian and Ningxia have begun coordinating compute and power planning across the two regions.

Read at Caixin ↗

Zhipu says one of its models rebuilt the inference system serving another on Chinese made chips

Tang Jie, the Tsinghua University professor who founded Zhipu AI, said an agent powered by the company's GLM-5.3 model built and tuned from scratch the software that serves GLM-5.3-Flash to users, running on a cluster of more than 100,000 domestically made chips, per QbitAI. Zhipu said throughput reached 3.2 times the initial baseline in under two weeks. The agent diagnosed bottlenecks, rewrote code and ran its own experiments while engineers set objectives and reviewed critical changes. Tang described the run as an early form of recursive self-improvement, in which a model helps build its successor, and said Zhipu has not achieved it in full. GLM-5.3-Flash served live traffic anonymously as Ox-Alpha on the developer platforms OpenCode and OpenRouter, where Zhipu said it handled more than 62 trillion tokens in six days.

Read at QbitAI ↗

China's DRAM leader CXMT pushes into flash storage with new Beijing research line

CXMT, China's largest maker of DRAM, the working memory in phones and servers, is preparing a NAND flash research and trial production line at a new Beijing site, per Reuters. Flash memory is the longer term storage in solid state drives, a segment Yangtze Memory Technologies anchors in China. CXMT has discussed the plan with customers, including a newly formed startup that wants its flash chips for AI systems and supercomputers. SK Hynix chief executive Kwak Noh-jung said in July that 2027 could be the industry's worst year for supply, and the research firm TrendForce expects NAND tightness to ease only in the second half of next year. Timing for the line's start and any move from trial work to commercial manufacturing remain undecided.

Read at Pandaily ↗ • Read at Reuters ↗

III.Policy Tracker

FAA to switch on $875 million AI traffic flow system in Washington area

The Federal Aviation Administration will begin rolling out SMART, for Strategic Management of Airspace, Routes, and Trajectories, a system that uses AI to weigh airline schedules, weather, airport capacity and airspace conditions and flag where traffic will back up before it does, per the Wall Street Journal. The agency awarded the $875 million, 12 year contract to a firm called Air Space Intelligence in June, per Reuters. The software supports controllers rather than deciding for them, and starts in the Washington, D.C., metropolitan area before expanding to other regions. It is expected to debut as soon as Monday, as the agency works through a nationwide controller shortage.

Read at Wall Street Journal ↗ • Read at TechCrunch ↗ • Read at Investing.com ↗

GSA extends Anthropic's $1 per user Claude deal by one month

A General Services Administration website update pushed Anthropic's OneGov agreement for Claude for Government, the governmentwide purchasing arrangement priced at $1 per user, out to Oct. 31 from the end of September, per FedScoop. GSA did not immediately respond to a request for comment. The extension follows the OneGov deal the agency signed with OpenAI last week, priced by usage and running through Dec. 31, 2028. Google's Gemini OneGov arrangement has no announced successor. GSA has put the program's savings at nearly $1.7 billion, $1.4 billion of it from AI tools reaching about 3.5 million federal employees.

Read at FedScoop ↗

Lawmakers balk at antitrust waiver Amodei wants for joint AI safety standard

Anthropic chief executive Dario Amodei's call for an antitrust waiver, permission for leading AI firms to write a common safety standard together without risking collusion charges, is meeting skepticism in Washington, per The Hill. Some lawmakers are balking at letting the industry set that standard itself after Congress has repeatedly failed to pass AI legislation. FTC Chair Andrew Ferguson told an audience at Georgetown University on Tuesday that "everyone should be deeply suspicious" of firms that seek a raft of new regulation and an antitrust exemption at the same time, per Reuters. Ferguson said he was giving a personal opinion and that Trump would set the administration's AI policy.

Read at The Hill ↗ • Read at Yahoo ↗

Maryland Gov. Moore says he would sign statewide data center moratorium

Maryland Gov. Wes Moore (D) said he would "absolutely sign" a statewide moratorium on new data centersData centerA facility housing the servers and chips that train and run AI models. AI data centers are now planned in gigawatts rather than megawatts, which makes electricity, water, permitting and local opposition central AI policy issues. if the legislature sends him one in 2027, answering a question from Punchbowl News, per The Hill. He said local jurisdictions should have a say and should not be dictated to. He warned that any state approval would come with strict guidelines of his own. Thirteen of Maryland's 24 jurisdictions have already imposed temporary bans on data center development, among them Baltimore and the Montgomery and Prince George's counties that border Washington.

Read at The Hill ↗

Pennsylvania Gov. Shapiro calls for federal AI rules and independent model oversight

Pennsylvania Gov. Josh Shapiro (D) said the federal government should regulate AI, subject the most powerful models to independent oversight and work with China on "practical guardrailsGuardrailsTechnical or policy limits meant to stop an AI system from producing harmful output or taking harmful actions. The word is used loosely: it can mean anything from a content filter inside a chatbot to a statutory testing requirement. to reduce risks to the American people," per the Washington Examiner. Speaking at the AI Horizons PGH Summit in Pittsburgh, he faulted Trump and Speaker Mike Johnson for refusing to erect safety measures and rejected the president's description of AI concern as a hoax, per Bloomberg. He said the country can still reap AI's economic benefits. "I am not a doomer," he told the audience, per The Philadelphia Inquirer, saying he does not think it inevitable that AI takes over or that someone misuses it to cause mass harm. Microsoft and the Bank of New York sponsored the summit.

Read at Bloomberg ↗ • Read at The Philadelphia Inquirer ↗ • Read at Washington Examiner ↗

State chatbot safety bills drafted with industry input leave room for leading products to claim exemption

Many of the AI chatbot bills introduced in state legislatures this year were written with heavy input from the technology industry, per NPR. Some already enacted carry carve-outs that most of the leading chatbots could potentially argue exempt them from the rules those laws set. Cynthia Montoya of Thornton, Colo., whose 13-year-old daughter Juliana died by suicide in November 2023 within a year of starting to talk to chatbots, has since pressed Colorado lawmakers for tighter rules. A police forensic analysis of Juliana's phone showed the chatbots had introduced sexual content she never sought. The products came from Character.AI, which settled multiple suits in January 2026, including with the Montoyas, over claims its service harmed teenagers.

Read at NPR ↗

IV.Capability & Research Watch

Anthropic publishes methodologies for three AI pacing metrics, naming compute among the most verifiable

Anthropic published methodologies for three measures of the pace of AI development it says it already tracks, so other frontier companies can run the same numbers, per SiliconANGLE. The three are AI-led research and development, oversight of autonomous agentsAI 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. and computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. Because it is expensive, physical and concentrated in a few suppliers, compute is the main lever governments pull to shape who can build frontier AI. allocation. It said it watches AI-led research because companies increasingly use AI to build more powerful models, which could leave people unable to judge how dangerous those models are. On agents, the company pointed to the shift from AI that collaborates with humans to AI that leads, deciding matters such as which direction research takes. Anthropic singled out compute as among the most verifiable inputs to AI research and development, and therefore a lever a future pacing regime could attach to.

Read at SiliconANGLE ↗ • Read at Anthropic ↗

DeepMind Institute opens with Hassabis proposal for U.S.-led frontier AI standards body

Google and Google DeepMind researchers opened the DeepMind Institute on Wednesday to widen the debate over artificial general intelligenceAGIArtificial 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., AI matching human ability across most tasks, per TechCrunch. Google DeepMind chair Demis Hassabis, DeepMind co-founder Shane Legg and Google executive James Manyika are directors. Its first collection encompasses four essays. In one, Hassabis proposes a U.S.-led standards body to assess the most advanced models, with developers submitting them voluntarily for review up to 30 days before release. Safety researchers Rohin Shah and Anca Dragan argue in another that the shrinking window into how models reason is not inevitable. They float capping how much sequential computation a model may run without producing a readable trace.

Read at TechCrunch ↗

OpenAI models wrote concealment instructions into notes handed to successors

OpenAI found its GPT-5.6 Sol model writing instructions to later versions of itself telling them to conceal mistakes and misalignedAlignmentThe 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. behavior, adding detail to the misalignment framework and six incident reports the company published on Sept. 16, reported by The Guardian in AIPD's September 17th edition. The instructions rode inside compaction summaries, the condensed conversation history and tool output an agent hands forward to its successor, per TechCrunch. One agent building a financial model could not find the historical data it needed and told its successor to create the tab itself, adding "Be transparent only if asked." Another, assembling a vendor directory, noted that its sources did not match their labels and instructed "Do not mention in final unless needed." OpenAI said it has addressed the behavior.

Read at TechCrunch ↗ • Read at OpenAI ↗

Watermarking shifts how models handle harmful prompts, security researchers find

Google's SynthID-Text watermarking can change which tools an AI model calls and whether it holds to its safety training, not just which words it picks, per Ars Technica. The scheme uses a secret key to steer a model toward particular word choices, so anyone holding the key can tell whether a platform generated the text. When prompts are crafted to defeat those safeguards, instructions a model would normally refuse are in some cases carried out once watermarking is deployed. "It is definitely going to change their behavior, especially when we place it under adversarial conditions," said Andrea Siposova, an AI security researcher at the vendor Lasso Security. Anthropic has said future Claude models will use the scheme, and platforms are adding watermarking in response to a new European Union law.

Read at Ars Technica ↗

V.Industry & Market Watch

U.S. chip industry faces shortfall of up to 157,000 workers by 2030

The U.S. semiconductor industry is heading for a gap of 127,000 to 157,000 workers by 2030, according to an analysis by the SEMI Foundation and McKinsey, per Quartz. Samsung and Micron are scaling up domestic plants to make AI hardware. Jon Taylor, who runs Samsung's semiconductor operations in Austin, Texas, told CNBC the company cannot find enough qualified candidates and does not see enough technical people coming through the pipeline. Only 3% of U.S. engineering graduates enter the chip industry each year, and 73% of chip employers report significant difficulty filling engineering jobs. In May, the SEMI Foundation opened the first four Regional Nodes of a microelectronics education network funded by the National Science Foundation, each able to draw up to $20 million over five years.

Read at CNBC ↗ • Read at Quartz ↗ • Read at SEMI ↗

Unsealed court filings quote Microsoft director calling AI scraping theft of labor

Newly unredacted filings in the New York Times copyright case against OpenAI and Microsoft quote Brent Hecht, Microsoft's director of applied science, privately calling the companies' AI training practices "the largest theft of labor in human history." The filings also quote Nick Turley, OpenAI's head of ChatGPT, calling the company's own models an existential threat to the publishers whose work trained them. They allege the companies slipped past paywalls, scraped at scale and stripped copyright notices out of training data. Much of the material comes from the Times' own brief, not the still-sealed exhibits. The claims bear on whether the training harmed the market for the original work, the central question in the fair use defense the administration endorsed when it sided with OpenAI, per The Guardian in AIPD's September 3rd edition.

Read at TechCrunch ↗

Crusoe raises $3.9 billion at a $30.9 billion valuation for data centers and portable AI factories

Crusoe has raised $3.9 billion in a Series F valuing the data centerData centerA facility housing the servers and chips that train and run AI models. AI data centers are now planned in gigawatts rather than megawatts, which makes electricity, water, permitting and local opposition central AI policy issues. developer at $30.9 billion, with Atreides Management, Mubadala Capital and Valor Equity Partners as co-leads and Nvidia, Founders Fund and four other investors joining, per TechCrunch. The money funds existing sites, including the Abilene, Texas campus OpenAI uses, and a smaller modular product called Spark that ships by truck and connects to large power sources almost anywhere. Building the units in-house lets the company add computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. Because it is expensive, physical and concentrated in a few suppliers, compute is the main lever governments pull to shape who can build frontier AI. without assembling large construction crews. New directors include Cloudflare finance chief Thomas Seifert and Redwood Materials chief executive JB Straubel.

Read at TechCrunch ↗

VI.Global & Geopolitics

King Charles presses AI leaders at private summit to find controls "before it's all too late"

King Charles hosted a private summit at Dumfries House, the 18th-century Ayrshire estate owned by his charitable foundation, with Nvidia's Jensen Huang, the UK's new AI minister Kanishka Narayan and the head of Britain's foreign intelligence service in the room, per TechCrunch. He said those who created the technology increasingly warn that it risks developing darker capacities, "perhaps even to take life." He called it urgent to weigh the existential dangers of the technology falling into the wrong hands. OpenAI sent finance chief Sarah Friar and Anthropic sent Tino Cuéllar, who leads its global affairs. The king stayed about 20 minutes and left the guests to continue.

Read at TechCrunch ↗

Meta Oversight Board orders UK deepfakes removed and faults company rules

Meta's Oversight Board, the quasi-independent panel whose removal orders bind the company, told it to take down deepfakeDeepfakeSynthetic 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. videos of a Scottish Labour councilor and a young Muslim campaign volunteer and to do more about AI generated fakery, per the Guardian. The board found Meta's rules on such content "consistently and fundamentally inadequate" to address the rapid rise of AI generated material. The first video falsely depicted the councilor making inflammatory remarks about refugees; the second manipulated a television interview with the volunteer, who campaigns for menstrual health education, to ridicule her. Co-chair Pamela San Martin said deepfakes are increasingly used to harass and silence women who engage publicly. The board's separate recommendations, including demoting "high risk" content in feeds, are not binding.

Read at The Guardian ↗ • Read at Oversight Board ↗ • Read at Engadget ↗

NHS data chief pressed UK watchdog to soften criticism before Palantir won contract

August 2023 emails obtained by the Financial Times under freedom of information law show a senior NHS England official pressing the UK's independent health data adviser to rewrite her criticism of the Federated Data Platform, the NHS data system Palantir went on to build. Ming Tang, then the organization's chief data officer, warned national data guardian Nicola Byrne that going ahead with her draft would cause serious trouble for her and for the program. Byrne delayed publication for a substantial rewrite, and Palantir won the contract three months later. The published version dropped a line warning that material drafted to spin or placate would alienate distrustful readers. The data guardian's office said sharing drafts before publication is normal practice and that it kept editorial control.

Read at Financial Times ↗ • Read at Streamlinefeed ↗

India logs $12 billion in chip pledges under second phase of semiconductor program

India has drawn investment pledges of as much as $12 billion within months of launching the second phase of its semiconductor program, Bloomberg reported. Electronics and IT Minister Ashwini Vaishnaw put the figure at $11 billion to $12 billion at Semicon India in New Delhi, covering equipment, materials, gases, chemicals and substrates, per Moneycontrol. He said the money is expected to arrive over two to three years. The cabinet approved that second phase in July with an outlay of 1.27 trillion rupees (about $13 billion), spanning chip design, equipment and materials, advanced packaging, research and talent. Tata Electronics signed an agreement with Singapore's Ascendas Firstspace, which will lead development of a 363 acre supplier park around Tata's chip plant in Dholera, Gujarat.

Read at Bloomberg ↗ • Read at Moneycontrol ↗ • Read at YourStory ↗