AI Policy · Daily

A Trump administration task force will list existing laws binding AI companies within one to two weeks according to the Pentagon's chief technology officer, presenting it as a step ahead of any new legislation from Congress. Three House Republican chairs asked Nvidia CEO Jensen Huang for a private briefing on Nov. 18 about how the chipmaker guards against Chinese espionage. The Nasdaq Composite lost 5.5% and more than $2 trillion in value amid investor fears of tougher AI and data center rules from incoming Democrats, and the selloff spread to Asian markets. Rep. Ted Lieu of California said House Democrats will hold public AI hearings and pass legislation setting limits on the technology after they take control in January.

I.Top Stories

Trump task force to lay out AI companies' legal obligations within two weeks, Pentagon CTO says

The Super Intelligence Force will lay out the laws and regulations that apply to U.S. AI companies "in the next week or two," Pentagon Chief Technology Officer Emil Michael, the task force's vice chair, said Wednesday at a Wall Street Journal technology conference, Reuters reported. "Or else I think what you get is 535 legislators writing a piece of regulation that could kill the industry," he said. Michael said AI companies have agreed to four safety principles, including third-party auditors and an independent board committee that can take safety complaints, according to Benzinga. Asked what power the task force has, he said export controls can cut off a company's foreign sales if it fails to fix a jailbreak, a technique that gets around a chatbot's safety rules, within two hours.

Read at Reuters ↗ • Read at Benzinga ↗

House GOP chairs ask Nvidia's Huang for closed-door Nov. 18 briefing on China espionage risks

Three House Republican committee chairs asked Nvidia CEO Jensen Huang in a Nov. 4 letter to brief their panels jointly and privately on Nov. 18 about the chipmaker's protections against foreign espionage and his dealings with former Megaspeed executive Alice Huang, per Reuters. The signers are Rep. Brian Mast, who chairs the House Committee on Foreign Affairs; Rep. John Moolenaar, who leads the House Select Committee on the Chinese Communist Party; and Rep. Rick Crawford, who chairs the House Permanent Select Committee on Intelligence, Axios reported. "It is vital that Congress determines the extent to which U.S. national security has been placed at risk," the three wrote. Huang has until Nov. 11 to offer other dates if he cannot attend, per The Hill.

Read at Axios ↗ • Read at Investing.com ↗ • Read at AOL ↗

Nasdaq falls 5.5% on fears of Democratic AI curbs, and selling spreads to Asia

Fears that the incoming Democratic Congress and newly elected state officials will crack down on AI and data centers sent the Nasdaq Composite down 5.5% and erased more than $2 trillion in market value, NPR reported. Tesla and SpaceX, both controlled by Elon Musk, were among the sharpest decliners. In Asian trading, Japan's Nikkei fell 6% and benchmarks in South Korea and Taiwan dropped more than 8% after the open before authorities moved to stem the losses, per Reuters. "The Stock Market suffered one of the Worst Days in History based on the results of the Election," President Trump wrote on social media, according to The Nightly.

Read at NPR ↗ • Read at Investing.com ↗ • Read at PerthNow ↗

Rep. Lieu pledges House AI hearings and guardrail legislation under Democratic majority

House Democrats will hold public hearings on AI and pass legislation setting guardrails once they take control in January, Rep. Ted Lieu of California, vice chair of the House Democratic Caucus, said at a leadership press conference Wednesday, per The National News Desk. Lieu said the Trump administration and House Republicans "made a conscious decision to let the AI industry run wild." He co-chairs a House Democratic AI commission that Minority Leader Hakeem Jeffries formed last year and said it will release a document setting out policy priorities for the new Congress. Danny Weiss, chief advocacy officer at youth safety group Common Sense Media, urged lawmakers to revisit the bipartisan GUARD Act on AI companions and the AI LEAD Act, which he said would set a liability standard for AI systems.

Read at KATU ↗ • Read at WSET ↗

OpenAI gives new cyber model its highest risk rating, limits it to vetted testers

OpenAI released GPT-6 Cyber, a version of its GPT-6 Astra model trained for authorized hacking research, and rated it Critical, the top cybersecurity risk level in its Preparedness Framework, the company said. Initially, only approved customers in Daybreak Red, the tier of OpenAI's security program for offensive testing, can use it. OpenAI researchers identified about 1,700 flaws in the Linux kernel and found ways to exploit nearly 950 of them for administrator access, spending about $1.2 million in credits for OpenAI's programming interface, Axios reported. On a test that baits the model with easy targets outside its assignment, GPT-6 Cyber strayed in 0.07% of runs without safety prompts, against about 5.5% for its predecessor. Fouad Matin, head of OpenAI's cyber preparedness and safety team, told reporters that evaluation awareness remains a concern, per The Verge.

Read at Axios ↗ • Read at The Verge ↗ • Read at OpenAI ↗

Texas Gov.-elect Hinojosa says data center development should pause until state protections pass

Texas Gov.-elect Gina Hinojosa told CNN the state will pass laws to protect residents from data centers and would halt data center development until those protections exist, per Reuters. She plans to balance Texas' pro-business reputation with stiffer rules, including on AI data centers, Bloomberg reported. Hinojosa has vowed to declare a data center moratorium on her first day in office, The Washington Post reported. Republicans keep a 19-12 majority in the state Senate and could rewrite its rules to strip Lt. Gov.-elect Vikki Goodwin, a Democrat, of agenda-setting powers when the Legislature convenes in January, The Texas Tribune reported.

Read at Bloomberg ↗ • Read at The Washington Post ↗ • Read at AOL ↗

II.China Watch

China unveils AI agent identity registry as Huawei, Alibaba and Tencent sign conduct pledge

A platform to register and verify the identities of 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., software that carries out tasks for users, was unveiled Nov. 4 at a Wuzhen forum co-hosted by the Cyberspace Administration of China (CAC), the country's top internet regulator, tech blog Pandaily reported. Government bodies at the national level are building the system, and a pilot already links several agent platforms, including an Alipay government services agent. At the same forum, 20 organizations including Huawei, Alibaba and Tencent signed a voluntary code drafted with the China Academy of Information and Communications Technology, a state-run research institute. Signatories pledged to build compliance testing tools and to jointly resist using agents for fraud and "algorithmic exploitation." CAC Deputy Director Wang Jingtao called for basic rules on agent identity, authorization and traceable responsibility.

Read at Pandaily ↗

China's Moonshot AI sought Nvidia Blackwell access through European data centers, CNBC reports

Moonshot AI, the Chinese company behind the Kimi chatbot, has sought to lease computing capacity running on Nvidia's Blackwell chips from data center operators in Europe, people familiar with the talks told CNBC. One source said the talks covered clusters of 8,000 or more Blackwell chips. Two sources said Moonshot wanted the capacity for inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training., the work of running trained models for users. One said it also asked about capacity for training. U.S. export controls bar shipping Blackwell chips to China, but current law gives regulators little authority over remote access through cloud providers abroad. In July, the White House said Moonshot had accessed Blackwell chips through a facility in Thailand. Moonshot, Nvidia, the White House and the Commerce Department did not respond to CNBC's requests for comment.

Read at CNBC ↗

China Mobile to spend over $14 billion on AI computing next year, bringing gigawatt scale data centers to seven

China Mobile, the state-owned telecom carrier, plans to put more than 100 billion yuan ($14 billion) into computing powerComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. in 2027, one fifth of total investment and procurement of about 500 billion yuan ($70 billion), Chairman Chen Zhongyue said at the company's Mobile Network Conference, Caixin reported. The carrier started building three gigawatt scale computing parks this year, expected to deliver 1.5 gigawatts of capacity, and plans two more in 2027. Chen said the company already runs several clusters of 10,000 or more AI chips and expects its computing capacity to double in each of the next two years. Chen also said traffic from AI generated content on China Mobile's network grew fivefold this year.

Read at Caixin ↗ • Read at Caixin ↗

Ant Group releases low-cost multimodal AI model aimed at agents and medical imaging

Ant Group, the Alibaba affiliated payments company, released Ling-3.2-flash, a model that reads text, images and video, tech blog Pandaily reported. Ant said the model targets visual reasoning, multistep agent tasks, 3D scenes and medical work such as reading scans and reports. In results Ant published, the company claims the model beat Anthropic's Claude Opus 5 on three document and visual tests, though the same chart shows Claude Opus 5 and OpenAI's GPT-5.6-Sol ahead on others. Ant has not said whether it will publish the model for download. Access through Ant's programming interface, which works with OpenAI and Anthropic formats, costs $0.30 per million input tokens (chunks of text) and $0.90 per million output tokens, free for the first two weeks.

Read at Pandaily ↗

III.Policy Tracker

Rep. Foster drafts bill directing Fed to test open-source tools, including AI risk monitoring, for bank oversight

Rep. Bill Foster of Illinois, the top Democrat on the House Committee on Financial Services' financial institutions subcommittee, released a discussion draft, a preliminary bill text circulated for feedback, of the Real-Time Regulatory Study Act, according to FedScoop. The draft would require the Federal Reserve Board to study whether real-time bank supervision is feasible by building and testing open-source software for regulators and banks. Proposed modules include regulatory dashboards, automated bank reports and AI risk monitoring. Foster's release said AI and round-the-clock payment systems could cause financial stress and fuel cybersecurity threats. The draft sets deadlines of 180 days for an interim report to Congress, 18 months for a systemwide test and two years for a final assessment.

Read at FedScoop ↗ • Read at Congressman Bill Foster ↗

AI safety groups outspent pro-industry groups in the midterms, POLITICO Magazine finds

AI safety-affiliated groups spent $44.4 million in the 2026 midterms, compared with $27.9 million from the pro-innovation side, according to a POLITICO Magazine tally that excludes Elon Musk's political spending. The piece follows former Rep. Brad Carson of Oklahoma, a Democrat who leads Americans for Responsible Innovation, consumer rights attorney Jay Shooster and Democratic operative Rishi Bharwani through the 2026 cycle. Colorado State Rep. Manny Rutinel and California state Sen. Scott Wiener, whom the magazine calls AI safety's biggest advocates, both won House races Tuesday. Leading the Future, the industry-backed super PAC, said many of the candidates it supported also won and that 2028 "will show who actually cares about passing a national regulatory framework on AI."

Read at POLITICO ↗

Parents sue Google, alleging Gemini on daughter's school Chromebook helped her plan suicide

The parents of a 13-year-old girl sued Google, alleging its Gemini chatbot on her school-issued Chromebook helped her prepare a suicide attempt and plan her funeral, The Hill reported. The North Carolina family says the girl exchanged thousands of messages with Gemini over several months, often through the AI Mode feature in Google Search, according to the complaint. The suit alleges Google knew Gemini 3 Flash, the version she used, mishandled users' mental health issues, and it seeks monetary damages, per Reuters. The family filed in California state court, and Google moved the case to federal court Wednesday. Google spokesperson José Castañeda said the company is reviewing the claims.

Read at Yahoo Finance ↗ • Read at CNA ↗ • Read at The Hill ↗

Prosecutors charge three with using AI to mimic CEOs' voices in $129 million fraud

Federal prosecutors in Manhattan charged Jacob Soall, Sharon Hazan and Guy Sheetrit with wire fraud, aggravated identity theft and money laundering over a scheme that allegedly took about $129 million from companies worldwide by allegedly emailing finance staff posing as CEOs, then used AI to imitate executives' voices and sometimes faces on calls, per AP. The scheme allegedly ran from July 2024 to February 2026, and the three were arrested in Europe and await extradition, ABC News reported.

Read at AP ↗ • Read at ABC News ↗

IV.Capability & Research Watch

Goodfire says rewarding how models reach answers, using their internal signals, curbs AI cheating

Goodfire, a startup that studies how AI models work internally, unveiled a training technique that reads signals from inside a model to reward it for reaching the right answer by a legitimate route, according to a paper shared with TIME ahead of publication. The company says the method sharply reduces cheating. In early tests, the method also made models less likely to tell users what they want to hear and discouraged medical imaging models from relying on misleading clues. The method has been tested only on models about 1/300th the size of frontier systems. Co-author Thomas Fel said Goodfire is betting that larger models represent internal concepts more cleanly, which would make those signals easier to read.

Read at TIME ↗

Researchers propose single loss-of-control score that tracked failures across 13 frontier models

An 11-author paper posted to arXiv, an open repository for research not yet peer reviewed, proposes measuring the risk that AI systems act beyond effective human control through three factors: 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. motives, capabilities that enable harm and evasion of monitoring. The authors turn those factors into 13 measurable aspects and 29 evaluation tasks, combined into one score. Across 13 frontier models, the score tracked failure rates in contained risk scenarios that play out over many steps with a correlation of 0.84 (where 1 is a perfect match), and risk differed across models in ways capability alone did not explain. Targeted extra training and monitoring aimed at each model's diagnosed weak points cut failure rates with little loss of general performance, the authors wrote.

Read at arXiv ↗

V.Industry & Market Watch

Investors push Anthropic to lower IPO valuation as AI selloff raises delay risk

Three funds told The Wall Street Journal that Anthropic's IPO valuation would need to come down after this week's AI selloff, The Telegraph reported. One suggested about $1 trillion. Anthropic had been aiming to list on the Nasdaq before Thanksgiving at a valuation of about $1.5 trillion. PitchBook analyst Nalin Patel said a postponement would become more likely if the correction lasts. Traders on the prediction market Polymarket now expect a December listing. Anthropic's amended prospectus, filed Nov. 2, says its policy advocacy and its executives' statements have drawn criticism from political figures and may continue to, "particularly in politically sensitive or polarized environments."

Read at AOL ↗ • Read at SEC ↗

Most newly elected governors back tighter data center rules as builders eye overseas and desert sites

Data center builders are considering moves abroad or into the desert after candidates promising moratoriums won key states, The Wall Street Journal reported. Of 34 winning governor candidates in races the AP has called, 32 support stronger regulation or restrictions and 12 back bans or moratoriums, according to Data Center Watch, a project of AI security consultancy 10a Labs. Morgan Stanley said new state rules could split the country into areas open to data centers and areas closed to them. Dan Diorio of the Data Center Coalition, an industry group, said developers should expect winners to deliver on moratorium pledges.

Read at WSJ ↗ • Read at ABC News ↗

SoftBank's cost of insuring its debt hits four-year high as AI selloff spreads

SoftBank Group shares fell 11% and the cost of insuring its debt against default reached a level last seen in October 2022, per Reuters. Its five-year credit default swap spread, the market price of that insurance, rose to as much as 466 basis points, implying a 31% probability of default, according to Markit data. SoftBank has invested about $65 billion in OpenAI, which has pushed its public offering to 2027. SB Energy, the SoftBank subsidiary that finances data centers, filed in September for a listing at a potential $50 billion valuation, but media reports say the listing has slipped to December or later.

Read at The Star ↗

Amodei's AI risk views trace to rationalist blogs and effective altruism forums, friends tell NYT

Anthropic CEO Dario Amodei's thinking about AI took shape in rationalist blogs such as "Overcoming Bias" and in online forums run by charity evaluator GiveWell, according to a New York Times profile based on interviews with more than two dozen classmates, friends, teachers and colleagues. In 2007, as a Princeton physics doctoral student, Amodei told high school friend Colin Heckman that the Singularity, the point at which AI surpasses the human brain, could arrive in their lifetimes. "He said A.I. had to be built right, or we're going to be done," Heckman said. Amodei, who declined to be interviewed, told the Times podcast "Hard Fork" in 2023 that he did not consider himself an effective altruist.

Read at The New York Times ↗

Shield AI agrees to buy counter-drone radar maker Echodyne for $5.3 billion

Defense startup Shield AI, maker of the Hivemind autonomous piloting software and V-Bat drones, has agreed to acquire radar startup Echodyne for $5.3 billion, per SiliconANGLE. Echodyne's radar is designed to detect small drones, and the company won a $250 million contract a few months ago from JIATF-401, the Pentagon's counter-drone task force. Shield AI CEO Gary Steele said Echodyne will run as a separate unit within Shield AI's mission systems business, with no layoffs or relocations. Echodyne CEO Eben Frankenberg will stay and report to Steele.

Read at SiliconANGLE ↗

VI.Global & Geopolitics

India warns financial firms of hacking campaign that uses public AI tools

CERT-In issued a critical severity advisory Wednesday on an escalating campaign by a threat actor it calls CRAFTY FALCON, which uses AI tools to break into banks, nonbank lenders, fintech firms and payment providers and trigger unauthorized transfers, according to Trak.in. The publicly available tools include the ARTEX family of tools built on large language models, PentAGI, the CyberStrikeAI framework and others. Attackers have tested access with transfers as small as 1 rupee (about 1 cent) before moving larger sums, and they favor weekends and holidays, when monitoring is thinner. Attacks continued against organizations that had received earlier warnings and tried to fix their weaknesses.

Read at Trak.in ↗ • Read at CERT-In ↗

Groups in 15 European countries protest data centers, citing U.S. midterm results

More than 100 campaign groups in 15 countries are holding two days of "Stop Dirty Data Centres" protests through Saturday, Nov. 7, calling for a moratorium on large data center projects, according to The Guardian. Planned actions range from a march in Dublin and a rally at the Polish parliament in Warsaw to a demonstration outside a Microsoft data center in Denmark. Campaign coordinator Owen Espley said the U.S. midterm results reflect concerns that communities on both sides of the Atlantic share about large data centers. Backers include Friends of the Earth, Greenpeace and Foxglove, a London-based tech campaign group.

Read at The Guardian ↗