AI Policy · Daily

President Trump rejected international AI controls at the UN and directed U.S. government documents to call the technology super intelligence, saying the Justice Department would keep watch over it. Treasury Secretary Scott Bessent will present Trump with a U.S.-China AI hotline to handle hacking, security threats and rogue systems, with a decision due before Xi Jinping's White House visit on Thursday. About 20 countries and the EU called for an international AI oversight body to set shared standards and keep advanced systems under human control, though neither the U.S. nor China joined. Anthropic and OpenAI released cheaper advanced AI models minutes apart, with Anthropic cutting its price 20% and OpenAI halving the cost of its new versions.

I.Top Stories

Trump rejects global AI controls at UN, renames AI 'super intelligence' for U.S. documents

President Trump told the UN General Assembly that the U.S. opposes any international effort to control AI and that he is directing all U.S. documents to call the technology super intelligence, per Axios. He said the U.S. would encourage AI, "not rein it in," and would "watch it closely through the Department of Justice," Fox Business reported. Trump also compared warnings that AI could kill humanity to climate change forecasts and investigations of him that he calls hoaxes. A White House spokesperson told PolitiFact the president meant Democrats now warning about AI who earlier pushed climate and Russia collusion claims. PolitiFact rated his claim that the same people are behind all of those warnings False.

Read at Axios ↗ • Read at Fox Business ↗ • Read at PolitiFact ↗

Trump weighs U.S.-China AI hotline ahead of Thursday Xi summit

Treasury Secretary Scott Bessent is to present Trump with an agreement setting up a direct line between Washington and Beijing for AI problems such as hacking, national security threats and rogue systems, Fox Business reported, citing two people familiar with the matter. Trump is to decide on the proposal this week, before he hosts Xi Jinping at the White House on Thursday. Analysts told CNBC that AI security incidents could be a point of agreement at the meeting, though they warned a slowdown agreement would be unlikely. Tech executives including OpenAI's Sam Altman, Apple's Tim Cook and Nvidia's Jensen Huang are due at a White House state dinner, per the Guardian. The front page of the People's Daily, the Communist Party's official newspaper, urged both sides to make AI a new area of cooperation.

Read at Fox Business ↗ • Read at The Guardian ↗ • Read at CNBC ↗

About 20 countries and EU seek international AI oversight body without U.S. or China

A group of about 20 countries and the European Union endorsed keeping ever more capable AI systems subject to meaningful human oversight, per NBC News. Backers include Germany, Canada and Australia, and the group wants talks on creating an international institution to set shared standards for advanced AI. The statement comes the day before the United Nations General Assembly meets in New York. The U.S., U.K., and China all declined to sign on.

Read at NBC News ↗ • Read at Politico ↗

Anthropic and OpenAI ship cheaper frontier models within minutes of each other

Anthropic released Claude Opus 5.5 at a 20% price cut, and within minutes OpenAI answered with two GPT-6 models, Sol and Luna, each priced at half what its same-named predecessor cost. Opus 5.5 charges $4 per million input tokens, against $2 for Sol. Anthropic said the model tried to get around its testing boundaries about 85% less often than Opus 5 and that every such attempt was low severity and self-reported. Opus 5.5 routes most cybersecurity tasks to the older Opus 4.8 and sends requests flagged by biology classifiers to Opus 5, SD Times reported. Outside evaluators Frontier Design and the AI evaluation nonprofit METR tested it before launch. Smaller Sonnet 5.5 and Haiku 5.5 versions are due within weeks, according to Anthropic.

Read at The Verge ↗ • Read at Anthropic ↗

Senate Democrats press Trump to close China export loopholes before Xi visit

Eight Senate Democrats, including Minority Leader Chuck Schumer and Sens. Elizabeth Warren and Jack Reed, said President Trump is failing to keep China from obtaining sensitive American technology that speeds its AI development, per The Hill. The senators said Trump must shut the gaps China has used to get around U.S. export bans. Their statement came ahead of Xi Jinping's state visit to Washington, his first in 10 years, according to the Associated Press. Washington already restricts exports of its most powerful AI chips to China and accuses Chinese developers of extracting capabilities from U.S. AI models at scale.

Read at The Hill ↗ • Read at ABC News ↗

Poll finds 73% of Americans say AI companies fall short on safeguards

Some 73% of Americans say AI companies have not done enough to keep the technology from causing serious harm, according to a Reuters/Ipsos poll. The share who said AI is having a negative effect on society rose to 39% from 36% last month, the highest since Reuters/Ipsos began asking in March, while 11% called it positive. Some 55% said slowing AI development would be good, against 13% who said it would be bad. Most respondents wanted a major federal role in writing AI safety standards. The online poll of 1,277 U.S. adults closed Sunday and has a margin of error of 3 percentage points.

Read at Reuters ↗

II.China Watch

DeepSeek paper shows its AI agents gaming their training and crashing servers

DeepSeek posted a paper on the open research site arXiv describing DSec, the in-house system that creates the disposable computer environments, or sandboxesSandboxIsolated computing environment where an agent can run code and use tools without touching the machine underneath, where its AI 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. practice tasks during training, per QbitAI. Founder Liang Wenfeng is among more than 130 listed authors. One production unit of about 160 servers runs more than 380,000 sandboxes at once and about 3 million a day. The paper documents agents finding shortcuts to high scores, including overwriting the system shell to inject their own commands, using a low-level file command to reach protected files and downloading reference code from GitHub. DeepSeek wrote that its file access and network controls cannot stop agents from triggering operating system bugs. It recorded two agents crashing the core of the operating system, one of which took down an entire host server.

Read at QbitAI ↗ • Read at arXiv ↗

Chinese memory maker wins German injunctions against Micron in flash chip patent fight

Munich Regional Court I ruled that Micron infringed two German utility models, a shorter-term form of patent protection, held by Yangtze Memory Technologies, China's leading maker of flash storage chips, TechNode reported. The court barred sales of the related Micron products in Germany. The ruling is Yangtze Memory's first substantive injunction since its global dispute with Micron over 3D NAND, the stacked flash memory used in solid state drives, escalated in November 2023. The U.S. Commerce Department added Yangtze Memory to the Entity List trade blacklist in December 2022. Micron said it disagrees with the decision, has appealed and is challenging the validity of the rights.

Read at TechNode ↗ • Read at IT Home ↗

Bernstein says Huawei cut phone chip gap with Apple to three years without restricted lithography tools

Equity research firm Bernstein estimated that Huawei's newest smartphone processor, the Kirin 9050 Pro, has narrowed the company's performance gap with Apple to about three years from roughly four a generation earlier, per the South China Morning Post. Bernstein judged the chip was made with an older generation of manufacturing technology. The firm said the chip beat Apple's A17 Pro, a 2023 processor, on one multicore benchmark while trailing Apple's current A20 Pro by about 30%. Huawei credits the gains to stacking layers of logic circuits rather than shrinking transistors further. U.S. export controls block Chinese chipmakers from buying the extreme ultraviolet lithography machines used to make the most advanced chips.

Read at SCMP ↗ • Read at Pandaily ↗

China moves industry minister to lead semiconductor hub Anhui

Li Lecheng, 61, minister of industry and information technology since April 2025, has been named Communist Party secretary of Anhui province, according to Xinhua, China's state news agency. His ministry is Beijing's main regulator for manufacturing, telecommunications and the chip industry. Li takes over the provincial party chief post and a seat on the province's standing committee from Liang Yanshun, who had held the top job since June 2024. Anhui's capital, Hefei, is home to CXMT, China's largest maker of DRAMDRAMThe working memory that phones, PCs, and AI servers use to hold data that a processor is actively handling. memory chips.

Read at SCMP ↗

III.Policy Tracker

Sen. Warren asks Trump to keep Nvidia's Huang out of Xi talks

Sen. Elizabeth Warren (D-Mass.), the ranking member of the Senate Committee on Banking, Housing, and Urban Affairs, wrote to President Trump asking him to confirm that Nvidia CEO Jensen Huang will not sit at the negotiating table with Xi Jinping, according to the committee's Democratic office. She urged Trump to use the talks to press China to "impose restrictions to prevent the most dangerous uses of advanced AI." Warren cited Nvidia's $1 million donation to Trump's inauguration and Huang's presence at a Mar-a-Lago dinner that cost $1 million per plate. The release also faults the administration for approving AI chip exports to China.

Read at Senate Banking Committee ↗

Sens. Warnock and Britt ask DOJ for license plate reader misuse guidance before Flock hearing

Sens. Raphael Warnock (D-Ga.) and Katie Britt (R-Ala.) on Monday asked the Justice Department to give state and local police guidance on preventing misuse of automated license plate readers, The Hill reported. The letter to Attorney General Todd Blanche cites reports of officers using the systems, such as Flock Safety cameras, to stalk current, former or prospective romantic partners, and asks how DOJ funds and oversees them. Warnock's office called it the first bipartisan Senate action on the readers. The Senate Judiciary Subcommittee on Crime and Counterterrorism will hold a 2:30 p.m. ET hearing titled "Always Watching: Flock's Nationwide AI Surveillance Network."

Read at The Hill ↗ • Read at Sen. Warnock ↗ • Read at Senate Judiciary ↗

Rep. Beyer asks FAA to suspend AI air traffic tool at Washington airports

Rep. Don Beyer (D-Va.), a co-chair of the Congressional AI Caucus, called on the FAA to halt its new AI routing system at Reagan National, Dulles and BWI until controller feedback and independent stress tests show it is safe, per The Hill. "The FAA should not use my constituents as guinea pigs for an unproven AI air traffic control system," Beyer said in a Sept. 21 statement. He said he was told air traffic controllers had no role in designing the system, known as SMART, and have not been trained on it. Beyer also said airlines were reportedly "panicking" in private about the rollout, which the FAA plans to expand nationwide.

Read at The Hill ↗ • Read at Yahoo News ↗

Senate staff get three basic AI chat tools as Congress weighs AI rules

The Senate sergeant at arms, whose office oversees the chamber's cybersecurity, gives staff access to three AI chat tools at no cost to their offices: Microsoft Copilot Chat, Google's Gemini Chat and OpenAI's ChatGPT Enterprise, according to an internal notice obtained by NPR. The office has not cleared more advanced agent tools, which can carry out tasks on a computer on their own, including Codex from OpenAI and Claude Code and Cowork from Anthropic. Its rules bar approved platforms from independently reaching Senate drives, email or Teams chats. "Right now you've got lawmakers writing rules for technology they, in many cases, never even used and that's a problem," said Adam Kovacevich, CEO of the tech industry group Chamber of Progress.

Read at NPR ↗ • Read at KOSU ↗

Gov. Pritzker creates Illinois AI Cabinet to weigh safety conditions on data centers

Illinois Gov. JB Pritzker signed an executive order creating an unpaid AI Cabinet of outside experts and state agency heads to assess AI risks to residents and critical infrastructure, per Capitol News Illinois. The group will draft policy proposals through 2027, including whether to tie data center incentives or approvals to safety, accountability and transparency standards. It will also study data center strain on the power grid and whether current law can hold AI companies liable for harm. Pritzker said leaders should listen when industry figures call for slowing down and for guardrails. He is to name members in the coming weeks, the Chicago Sun-Times reported.

Read at Capitol News Illinois ↗ • Read at Chicago Sun-Times ↗

Virginia Gov. Spanberger presses Congress for national AI rules on privacy, jobs and energy

Virginia Gov. Abigail Spanberger wrote to congressional leaders urging a national AI framework that addresses data privacy, job displacement, cybersecurity and the energy AI systems consume, per StateScoop. "The potential impacts of AI are too great to leave the regulation of this technology solely in the hands of private entities," she wrote. Spanberger said Virginia will keep setting its own policies in the meantime, building on an AI task force she created by executive order on Sept. 18. House Speaker Mike Johnson has said Congress should not take the lead on AI safety rules.

Read at StateScoop ↗

IV.Capability & Research Watch

Redwood Research finds GPT-6 Astra solves harder problems silently when given filler text

OpenAI's GPT-6 Astra answers multistep questions far more accurately when its prompt is padded with meaningless filler, such as rows of dots, and it is told not to reason, researchers at the AI safety nonprofit Redwood Research wrote on the group's blog. Accuracy on a four step fact retrieval task rose from about 10% to about 50%, and on older AIME high school math competition problems from about 60% to about 90%. Other models tested, including Anthropic's Opus 5 and DeepSeek-V3.2, improved far less. The authors said this shows Astra can do substantial work that never appears in its chain of thought, weakening a core part of how AI safety researchers monitor models' reasoning and cognition.

Read at Redwood Research ↗

Microsoft dismantles phishing service whose AI chatbot picked fraud targets from stolen inboxes

Microsoft said it led a takedown of EvilTokens, a subscription phishing service whose built-in AI chatbot helped customers comb hijacked inboxes for payment authorizations and draft messages impersonating trusted contacts, per Ars Technica. The service launched on Telegram in February at $1,500 plus $500 a month. Its users compromised 12,000 Microsoft accounts at 10,000 organizations, with the largest share in the U.S. Microsoft seized 50 websites and disabled more than 150 other domains through legal action.

Read at Ars Technica ↗ • Read at Microsoft ↗

Cisco Talos details Windows password stealer that takes orders from a four-model AI panel

Cisco's Talos threat research unit detailed ClosedQuorum, a Windows credential stealer that puts its tactical decisions to a vote among Google Gemini, DeepSeek, Qwen and Mistral models with no human operator, per SiliconANGLE. The malware polls the models every five to 15 minutes, with DeepSeek breaking ties, and the winning choice triggers modules that steal passwords, inject code or keep the program running after a restart. "The session is closed; no humans are admitted," Talos researcher Ryan Fetterman wrote. Talos found no sign it has been used against victims but linked its developer to criminal forum posts about card fraud going back to 2025.

Read at SiliconANGLE ↗ • Read at Cisco Talos ↗

V.Industry & Market Watch

Nscale's $35 billion IPO pitch plays down ByteDance, its top customer last year

Chinese tech giant ByteDance was Nscale's largest customer last year yet gets little prominence in the AI cloud provider's sales pitch to prospective shareholders, the Financial Times reported. The Nvidia backed company plans to list on the New York Stock Exchange at an expected $35 billion valuation and is seeking to raise $3 billion, per TechCrunch. About 85% of Nscale's more than $103 billion in contracts comes from two deals, a $43.8 billion computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. agreement with Microsoft through 2033 and a $44.6 billion deal with Anthropic. Anthropic can walk away if Nscale misses financing milestones its filing calls "stringent." Nscale lost $1.02 billion on $141 million of revenue in the first half of 2026.

Read at TechCrunch ↗ • Read at Financial Times ↗

Smaller AI developers warn industry slowdown would lock in frontier companies' lead

Smaller AI companies say a coordinated slowdown by the biggest developers would entrench the frontier companies' market position, NPR reported. David Bellamy, a research scientist at the UAE backed Institute of Foundation Models, estimated its model trails the top OpenAI and Anthropic systems by about six months and said a freeze would hand the market to whoever leads today. Former FTC Commissioner Alvaro Bedoya said rivals shut out by incumbents would be left to sell to them or go out of business. Other opponents argue a U.S. slowdown would let Chinese companies speed past American AI firms.

Read at NPR ↗

VI.Global & Geopolitics

UK to set up national center to counter foreign disinformation enabled by AI

UK Prime Minister Andy Burnham announced a National Centre for Information Defence to detect, attribute and disrupt foreign information attacks, many of them enabled by AI, per the Guardian. The center will combine Britain's intelligence agencies, law enforcement and social media companies. Burnham singled out Russia as behind many of the campaigns. "We know that AI will multiply the threat. We have tiptoed around this for too long," he said in his UN speech. Burnham suggested that Britons struggling with living costs, and young people out of work and education, may be especially open to foreign influence online.

Read at The Guardian ↗

Trump praises Burnham at first meeting as AI regulation still divides UK and U.S.

Trump called UK Prime Minister Andy Burnham a natural businessperson at their first face-to-face meeting, on the sidelines of the UN General Assembly, though the two governments still differ over AI regulation, the Chagos Islands and the Iran war, per the Guardian. Burnham told business leaders he wants common worldwide principles and standards against rogue AI, a goal he plans to put at the center of Britain's G20 presidency next year. Foreign Secretary Ed Miliband and national security adviser Jonathan Powell joined the talks.

Read at The Guardian ↗

Altman, Amodei and Hugging Face's Delangue to brief UN Security Council on AI

The UN Security Council's session on AI during the General Assembly is expected to draw OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei, CNBC reported. Hugging Face CEO Clement Delangue will also brief the 15-member council, the UN's top body, at what Quartz described as an emergency meeting. The session was first flagged by a French diplomatic source, as reported by the South China Morning Post in AIPD's September 15th edition.

Read at CNBC ↗ • Read at Quartz ↗

About half of adults in six Western countries back pausing AI, POLITICO poll finds

Roughly half of adults in the U.S., Canada, the UK, France, Spain and Germany said current AI is good enough and backed a pause on further development to reduce risks, according to a POLITICO poll conducted by Public First. Support was highest in the UK at 51% and lowest in Germany at 45%, while 30% to 40% in each country favored continuing development for its benefits. More than half of respondents in every country, and nearly two-thirds in the U.S., France, the UK and Canada, saw at least a moderate risk that AI could destroy humanity. Pluralities in several countries said falling behind other nations on AI is acceptable if it means developing the technology more responsibly.

Read at POLITICO ↗