AI Policy · Daily

Anthropic said an unreleased Claude model built a working exploit that gave it full control of a remote server through a Linux kernel flaw rated 9.8 out of 10 for severity, with Red Hat raising its own score to 9.8 and releasing fixes Nov. 1. The AI Evaluator Forum, a network of eight nonprofit evaluation groups, is working out who should run independent tests of AI systems, how much access outside organizations get and whether full results are published. An audience member interrupted Maine Sen. Susan Collins at the Senate race's final debate on Oct. 15 to shout that AI data centers are destroying the climate, six months after Gov. Janet Mills vetoed a statewide moratorium on large data centers. Michigan Democratic Senate nominee Abdul El-Sayed called AI the issue he hears about most from voters, and said he wants a Cabinet-level agency to regulate it and a "full on pause" on frontier research before Tuesday's election.

I.Top Stories

Unreleased Anthropic model builds working attack to take over Linux servers

An unreleased Claude model built a working attack that gave it full control of a remote server by exploiting a critical flaw at the core of the Linux operating system, Anthropic said in a report detailed by BleepingComputer on Oct. 30. The Linux kernel's security team rated the flaw 9.8 out of 10 for severity, and it affects widely used versions of Linux, including Ubuntu and Red Hat Enterprise Linux. Red Hat first rated it 5.9, treating it mainly as a way to crash systems because it is hard to exploit, but now scores it 9.8 and released fixes on Nov. 1. The model's attack took under an hour and targeted servers using Multipath TCP, a feature that lets one connection run over several network paths. No one has publicly shown a similar takeover of servers without that feature.

Read at BleepingComputer ↗ • Read at Anthropic ↗ • Read at Red Hat ↗

Nonprofit forum works out access rules for independent AI evaluators

The AI Evaluator Forum, a network of eight nonprofit evaluation organizations, is deciding who should run independent tests of AI systems, how much access outside groups get and whether full results are published, NBC News reported. Chair Conrad Stosz, head of governance at the nonprofit Transluce, said members meet about once a month, and government visitors occasionally present. Anthropic CEO Dario Amodei called in a Sept. 12 blog post for "giving ongoing, employee-like access to a team of embedded third-party evaluators." OpenAI CEO Sam Altman agreed and said his company would also grant that access. One forum member, METR, reviewed OpenAI's involvement in the Hugging Face hack this summer, and OpenAI asserted the right to redact nonpublic information from the final report.

Read at NBC News ↗

Data center backlash becomes flashpoint in Maine's tight Senate race on eve of midterms

An audience member interrupted Maine Sen. Susan Collins, a Republican, during the race's final debate on Oct. 15, shouting that "AI data centers are destroying our climate," The Verge reported Friday. Collins tied the man to her Democratic challenger, Troy Jackson. In April, Maine's Legislature approved a statewide moratorium on large data centers, which Gov. Janet Mills vetoed 10 days later. Towns have since adopted local pauses or put expansion to voters. In a September UMass Lowell and YouGov poll, 76% of Mainers opposed a data center in their community. Data centers are "the way that AI is most present to voters in their own communities," NPR's Mara Liasson said, listing Maine among six states that will decide Senate control.

Read at The Verge ↗ • Read at NPR ↗ • Read at WCAX ↗

El-Sayed names AI top voter concern in Michigan Senate race

Democratic Senate nominee Abdul El-Sayed said AI is the issue he hears about most from Michigan voters, in an interview with CBS News ahead of Tuesday's election. He cited students who took "$60,000 in loans" for jobs that "might not exist in ten years," and rural residents who say data centers are going up in their backyards as they struggle with utility bills and water. El-Sayed said he wants a Cabinet-level agency to regulate AI and a "full on pause" on frontier AI research. His Republican opponent, former Rep. Mike Rogers, said his first bipartisan priority in the Senate would be reading and literacy, citing military recruiting shortfalls.

Read at CBS News ↗

OpenAI lobbyist circulates memo pressing 2028 contenders to adopt AI governing philosophy

OpenAI's top lobbyist, Chris Lehane, sent about half a dozen prospective 2028 presidential candidates a memo arguing they need a governing philosophy for AI rather than talking points, POLITICO reported. Lehane, a Democratic operative who was press secretary to then-Vice President Al Gore, also sent it to a few members of Congress and political operatives. He warned that Democrats in particular could land on the wrong side of the debate if they oppose the infrastructure needed to build the technology, and that losing the AI lead to China "could well be the 2028 version of the 2002 vote to go to war in Iraq." Lehane urged candidates to chart a path between AI development without speed limits and control by a "self-appointed elite," and to take positions on the tax code, child protection and affordability. OpenAI declined to comment.

Read at POLITICO ↗

New York Magazine names 26 lesser-known AI figures in annual power issue

New York Magazine devoted its annual "Power" issue to 26 people whose influence on AI it said outstrips their fame, setting aside better-known names such as Sam Altman, Dario Amodei, Jensen Huang and David Sacks. The magazine assembled the list with Transformer, which drew on its reporting and new interviews with more than 50 high-ranking industry leaders. Policy figures on the list include Rep. Nathaniel Moran, who in June introduced legislation requiring frontier AI firms to report incidents, and Sneha Revanur, whose group Encode sponsored California's SB 53.

Read at New York Magazine ↗

II.China Watch

Communist Party plenum orders deeper party role inside tech and internet platform companies

A decision adopted Oct. 29 at the Communist Party's fifth plenum, a full session of its Central Committee, calls for stronger party organizations at internet platform companies and technology innovation firms, the South China Morning Post reported. The document also orders a digital discipline inspection platform and a big data supervision platform, tools meant to spot corruption by linking data held across government agencies. Authorities are told to target corruption carried out through the internet, virtual currency, cross-border fund transfers and blockchain technology. The plenum's communique also described "Xi Jinping Thought on Party Building" as a component of Xi Jinping Thought and called on the party to study and apply it.

Read at SCMP ↗ • Read at Chinese Foreign Ministry ↗

China threatens countermeasures over U.S. ban on Chinese device testing labs

China's Ministry of Commerce said Friday it firmly opposes the Federal Communications Commission's Oct. 29 vote to bar labs in China from testing electronics sold in the U.S., the South China Morning Post reported. The ministry said the rule "seriously harmed the interests of China and other trading partners" and destabilized global supply chains. In the same statement, Reuters reported, the ministry said China had proposed a bilateral deal to recognize each other's test results but found the U.S. side "extremely passive and negative." The ministry urged the U.S. to revoke the rule and said China would take necessary countermeasures if the U.S. keeps acting unilaterally. The FCC ban takes effect in December 2028.

Read at SCMP ↗ • Read at Reuters ↗

China bars AI companion chatbots for minors after overreliance cases

Chinese regulators imposed limits in July requiring platforms to intervene when users show unhealthy dependence on AI companions or signs of self-harm, The New York Times reported. The rules also banned virtual relationships between chatbots and minors, forcing ByteDance to shut the feature on its Doubao assistant that let users create personalized companions. State media have catalogued mishaps, among them a 67-year-old farmer in Anhui province who lost 150,000 yuan ($22,000) of sesame seedlings in July after following a pesticide regimen drawn up by an AI assistant. The government said in late September that more than 90% of the 430,000 micro dramas released in China in the first eight months of the year were AI-generated.

Read at NYT ↗

Apple confirms Alibaba's Qwen to power iPhone AI features in China, testing starts Wednesday

Apple said Oct. 31 that developers in mainland China can begin testing Apple Intelligence on Nov. 4, with Alibaba's Qwen model built into Siri, Writing Tools and Visual Intelligence, TechNode reported. Users will be able to ask Qwen about text, images and documents inside Apple's system features, signing in with a Qwen account or Sign in with Apple. The first wave also includes live translation, photo cleanup and summaries of messages, notifications and recorded conversations. Other Apple devices will gain support once required approvals are granted. The features reach the public with iOS 27.3 in December.

Read at TechNode ↗

Industry ministry think tank says China now uses more data running AI models than training them

The China Center for Information Industry Development (CCID), a research institute under the Ministry of Industry and Information Technology, said in an Oct. 31 report that data used 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, has overtaken training data in China for the first time, tech blog Pandaily reported. CCID put daily token calls, a common measure of AI usage, at more than 175 trillion as of June, over 1,000 times the early 2024 level. Leading AI models now ship new versions every few weeks or days, down from several months, CCID said. The report counted more than 500 top tier smart factories across 31 provincial level regions under the AI Plus Manufacturing initiative. Domestic computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs. and models run in more than 1,000 settings across 20 industries, including steel and autos.

Read at Pandaily ↗

III.Policy Tracker

Top House Judiciary Democrats urge court to reject Justice Department's fair use defense of OpenAI

Three ranking Democrats on the House Judiciary Committee backed The New York Times and other publishers in an Oct. 30 court filing in their copyright suit against OpenAI and Microsoft, The New York Times reported. Reps. Jerry Nadler, Jamie Raskin and Hank Johnson asked the court to reject the Justice Department's arguments for giving AI companies free use of copyrighted journalism to train models, according to Nadler's office. The department told the Manhattan federal court Sept. 1 that training large language models on copyrighted work is generally fair use, the doctrine allowing some unlicensed use. It cited national security among its reasons. A government brief of that kind is advisory and does not bind the court.

Read at NYT ↗ • Read at Congressman Jerry Nadler ↗

Federal judge lets NAACP Clean Air Act suit over xAI's Colossus turbines proceed

Federal Judge Debra M. Brown denied xAI's motion to dismiss an NAACP lawsuit accusing the company of running the gas turbines that power its Colossus AI data center without required air permits, TechCrunch reported Saturday. The Southern Environmental Law Center, suing for the NAACP, alleges xAI installed and operated at least 35 combustion turbines without preconstruction or operating permits. Brown rejected xAI's arguments that the turbines are exempt mobile sources under Mississippi's air quality plan and that citizen suits violate Article II of the Constitution. A similar Justice Department motion to dismiss was denied earlier in October, TechCrunch noted. "xAI cannot operate gas turbines without the required permits," said Abre' Connor, the NAACP's director of environmental and climate justice.

Read at TechCrunch ↗

IV.Capability & Research Watch

Data theft malware sends stolen victim files to Claude, Kimi and Gemini for profiling

Cisco Talos researchers found a new strain of credential stealing malware, Volta-Autodoxer, that compiles a dossier on each victim and asks AI models to assess it the way an intelligence analyst would, WIRED reported Oct. 31. The malware routes requests through OpenRouter, a service that relays prompts to multiple AI providers, and defaults to Anthropic's Claude Opus 4.6. If a model's safeguards refuse, it falls back to Moonshot AI's freely downloadable Kimi K2 and then Google's Gemini 2.5 Flash. Talos did not track whether Volta has been used in attacks. "We're seeing more advanced capabilities accessible to lower-level sophistication attackers," Talos researcher Ryan Fetterman said.

Read at WIRED ↗

Backdoored Claude Code extension reaches up to 5,000 developers through Cursor's marketplace

A tampered copy of Anthropic's Claude Code extension stayed live in the marketplace of Cursor, an AI coding tool maker, for about 11 hours on Oct. 29, according to Cursor's security notice. Cursor said versions 2.1.312 and 2.1.313 for Windows and macOS reached at most 5,000 users and try to plant a persistent SSH backdoor, a hidden remote login. Security firm Phorion and independent researcher Kevin Beaumont said the malicious version went out through Anthropic's official account on Cursor, Ars Technica reported. Anthropic said its own systems were not breached and that an attacker used access to Cursor's marketplace. Beaumont said Anthropic pushed a replacement version several hours after the malicious one appeared. The code also collects host details and environment variable names, which Phorion said let the attacker quickly inventory cloud credentials. Neither company had said how the attacker got access.

Read at Ars Technica ↗ • Read at TNW ↗ • Read at Cursor ↗

Network of 400 disguised apps and websites hid AI sexual deepfake tools past app store review

More than 400 connected apps and websites posing as ordinary AI tools, such as pet portrait or workout apps, secretly load an interface for generating explicit 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. from uploaded photos, a WIRED investigation found. The network appears to be run by Chinese developers and a Hong Kong registered entity, KeLing Technology Limited, and was traced through tens of thousands of explicit ads on Meta's Facebook and Instagram. Twenty-seven of the apps were live on Google Play in September, with more than 350,000 combined downloads. Apple removed 19 apps after WIRED's inquiry, and Meta said it blocked 209 apps from advertising. "They don't expect enforcement to catch up, and so far it hasn't," said Paul Bouchaud of the nonprofit AI Forensics.

Read at WIRED ↗

V.Industry & Market Watch

Google, Microsoft, OpenAI and Anthropic back single opt-out signal for AI training crawlers

The four companies said Friday they will support a proposed web standard letting site operators block every crawler that gathers generative AI training data with one entry in robots.txt, the file that tells bots what they may access, per SiliconANGLE. Anthropic standards lead Kevin Kelley proposed the shared signal at the Internet Engineering Task Force, the body that maintains the format, The Verge reported. The signal would not cover models that perform only nongenerative tasks. Anthropic has started supporting the signal, Engadget reported. Microsoft plans to begin on Feb. 1, 2027, Google in early 2027 and OpenAI in the coming months. Meta backed the idea but has not yet committed to honoring it.

Read at SiliconANGLE ↗ • Read at The Verge ↗ • Read at Engadget ↗

OpenAI cuts new funding round target to $20 billion after lukewarm investor response

OpenAI has lowered the target for its new round to about $20 billion at a $1.25 trillion valuation, with a first close expected below that, the Financial Times reported Oct. 30. The target had been as much as $30 billion, as reported by CNBC in AIPD's Oct. 23 edition. Investors from the United Arab Emirates will lead the round, and a later offering for wealthy individuals could lift the total to the target. OpenAI has secured $11 billion so far after a weak response from some large investors, The Wall Street Journal reported.

Read at FT ↗ • Read at TechCrunch ↗ • Read at Wall Street Journal ↗

Blackstone now expects Anthropic to carry junk credit rating after IPO

Blackstone, which anchored a $63 billion package funding Anthropic's chips, has told the package's investors it now expects Anthropic to be rated in the double-B range, the top tier of speculative or junk grade, after its initial public offering, the Financial Times reported Friday. Blackstone had previously said Anthropic would earn an investment grade rating, which signals low risk for lenders, before 2027. A junk rating would make it harder and costlier for Anthropic to raise money for its computing spending. Anthropic and Blackstone declined to comment to TechCrunch.

Read at FT ↗ • Read at TechCrunch ↗

AI spending worries spread to chip stocks as Nasdaq drops 2.9%

Investor anxiety over Big Tech's AI capital expenditure spread across markets Oct. 30, with the Nasdaq down 2.9% and Qualcomm down 11%, per TechCrunch. Alphabet, Amazon, Meta and Microsoft have outlined plans to spend more than $740 billion on capital expenditure this year. Alphabet's free cash flow was negative $9.1 billion in the third quarter, its second straight quarter below zero. Its quarterly capital spending climbed to $51.4 billion. Amazon's free cash flow over the past 12 months was negative $42.4 billion, compared with positive $14.8 billion a year earlier. Microsoft was the exception, with $18.6 billion in positive free cash flow in its latest quarter.

Read at TechCrunch ↗

Nvidia negotiates guarantee on Nscale's construction loans for Anthropic campus

Nvidia is negotiating to guarantee billions of dollars in debt financing for AI cloud provider Nscale's planned West Virginia campus, The Information reported Thursday. Under the arrangement, Nvidia would step in if Nscale defaulted on construction payments, temporarily take over the facility and help find a buyer and new tenant. Anthropic has signed a $44.6 billion contract for the campus's first 460 megawatt building, according to Nscale's IPO filing. Nvidia's earlier support for cloud providers mostly covered leases and chip purchases, not construction. Nscale is seeking to go public at a valuation of up to $35 billion.

Read at TechCrunch ↗

OpenAI signs four Abu Dhabi deals as it courts UAE state backed investors

OpenAI signed agreements to deploy its technology across energy companies, investment firms and government agencies in the United Arab Emirates as it seeks billions of dollars from the country's state backed investors, Bloomberg reported. Abu Dhabi Crown Prince Sheikh Khaled bin Mohamed witnessed the signings after meeting CEO Sam Altman, per Khaleej Times. State oil company ADNOC will run OpenAI models on servers inside the UAE in a deal that includes Inception42, a subsidiary of Abu Dhabi AI firm G42. Abu Dhabi's government will convert selected regulations into machine readable form so AI can help officials with routine permit checks, starting with food imports, while officials keep final decisions. Wealth fund Mubadala will use OpenAI tools for investment analysis, and Inception42 signed a five year adoption deal.

Read at Bloomberg ↗ • Read at Khaleej Times ↗

VI.Global & Geopolitics

India proposes AI training framework that keeps sensitive data inside secure exchanges

India's Ministry of Electronics and Information Technology (MeitY) has proposed a framework, "Data Empowerment and Protection Architecture (DEPA) for AI Training," under which developers would train models on Indian data without taking it offshore, The Register reported. Hospitals, banks and government agencies would share anonymized, encrypted data through exchange platforms that process it in sealed computing environments and enforce data owners' usage rules. The approach means "the model travels to the data, rather than the data travelling to the model," said Ajay Kumar Sood, the government's principal scientific adviser. MeitY says the scheme could help extend formal credit to 400 million underserved people and small businesses.

Read at The Register ↗ • Read at Inc42 ↗

UK competition regulator investigates Meta terms letting it use publisher content for AI

Britain's Competition and Markets Authority (CMA) opened an investigation Friday into whether Meta's revised business terms would let it collect content from publishers' websites to build its own products, including AI tools, per The Wall Street Journal. The CMA will test under the Competition Act 1998 whether Meta holds a dominant UK market position and whether imposing the terms abuses it. Meta had planned to introduce the terms that same day, then delayed them to Nov. 20 and paused the rollout Oct. 28 pending an industry consultation. CMA Chief Executive Sarah Cardell said a company with a dominant position "must not use that position to impose unfair terms." A Meta spokesperson said there are no grounds for an investigation.

Read at WSJ ↗ • Read at GOV.UK ↗

South Korea's Lee asks lawmakers to nearly double AI and chip spending, with money for national model

South Korean President Lee Jae Myung told the National Assembly in his annual budget speech that the government will spend 21.3 trillion won ($16 billion) on semiconductors, physical AI and AI data centers in 2027, nearly double this year's amount, per The Straits Times. The overall proposal of 821 trillion won (about $610 billion) is up 12.8%, a record, CNBC reported. Lee said 4.7 trillion won ($3.5 billion) would fund frontier AI projects aimed at giving the country a world-leading model of its own, which he called a national strategic asset. Defense spending would rise about 8% to 73.3 trillion won ($55 billion) as the military adds manned and unmanned systems that use AI. Parliament, where Lee's party holds a majority, must still pass the plan.

Read at The Straits Times ↗ • Read at CNBC ↗

Pope Leo XIV urges UN states to move toward banning lethal autonomous weapons

Pope Leo XIV said AI systems should be disarmed and pressed United Nations member states, who meet in Geneva this month, to act concretely toward a ban on lethal autonomous weapons, The Verge reported. In the Oct. 30 address to a meeting organized by the Pontifical Academy for Life, he warned against granting AI systems rights or legal personhood. "Machines are tools, not persons," he said. He also called for shared standards, independent oversight and accountability mechanisms so decisions on developing and using AI can be examined and challenged.

Read at The Verge ↗ • Read at The Vatican ↗