AI Policy · Daily

The Department of Homeland Security ordered civilian agencies to purge DeepSeek products and stop buying them, a ban that also covers the Chinese AI developer's affiliates and any products built on its technology. A federal judge denied release for a former OpenAI engineer accused of planting a recording device in a company conference room, after prosecutors said Saudi officials paid for his family's flight out of the country and offered to cover his bail. The State Department sanctioned a Chinese venture firm and its affiliates for seed funding a startup whose founder, a former Google engineer, was convicted in January of stealing the company's trade secrets. Senate Minority Leader Chuck Schumer asked the FCC to block AI political robocalls, opposing a conservative group's petition to loosen consent rules for such calls to cellphones before the midterms.

I.Top Stories

DHS bars DeepSeek from federal civilian IT under supply chain security law

The Department of Homeland Security has ordered civilian federal agencies to pull products from Chinese AI developer DeepSeek out of their systems and stop buying them, CyberScoop reported. The exclusion also reaches DeepSeek affiliates and third-party products built on the company's technology, according to SiliconANGLE. DHS invoked the 2018 Federal Acquisition Supply Chain Security Act after a review by the Federal Acquisition Security Council, an interagency panel on risky technology vendors. Homeland Security Secretary Markwayne Mullin signed the exclusion letter June 26, months before the order was posted to the Federal Register this week. Mullin wrote that DeepSeek data reaching China "could produce actionable intelligence on U.S. persons and/or government programs." Agencies wanting an exception must apply 60 days ahead and show a mission critical or national security need.

Read at CyberScoop ↗ • Read at SiliconANGLE ↗

Judge keeps ex-OpenAI engineer jailed after prosecutors cite Saudi backing

A federal judge in San Francisco turned down former OpenAI engineer Abdulrahman Alfozan's bid for release to home confinement, The Associated Press reported. Alfozan is charged with planting a recording device in an OpenAI conference room. In a Sunday filing, prosecutors said a Saudi military official booked and paid for the one-way flight his wife and son took out of the U.S. the day after his arrest. Saudi officials also offered to cover his bail, the New York Times reported. Judge Alex Tse likened that offer to a "blank check," per Wired. Prosecutors said Alfozan had "considerable support from a foreign government." They said he was in frequent contact with a Saudi royal who repeatedly tried to recruit him to an AI venture in the kingdom.

Read at AP ↗ • Read at The New York Times ↗ • Read at The Washington Post ↗ • Read at Wired ↗

State Department sanctions Lu Qi's MiraclePlus for seeding ex-Google engineer's startup

The State Department sanctioned MiraclePlus, a Chinese venture firm founded by former Microsoft executive Lu Qi, and three affiliates for giving seed money to Shanghai Zhisuan Technology, per Bloomberg. Zhisuan, which builds software to help train AI models, was also sanctioned. Its founder, former Google engineer Ding Linwei, was convicted in January on seven counts of trade secret theft. The five entities were designated under the Protecting American Intellectual Property Act, the South China Morning Post reported. The measures could include freezing their U.S. assets and restricting their banking, The Edge Malaysia reported. According to the State Department, Ding applied to MiraclePlus' incubator by promoting his knowledge of Google's technology. MiraclePlus grew out of Y Combinator's abandoned China program and has backed SiliconFlow, a platform that runs AI models for developers, and robot maker AgiBot.

Read at Bloomberg ↗ • Read at The Edge Malaysia ↗ • Read at South China Morning Post ↗

Schumer presses FCC to deny Club for Growth bid to ease AI robocall limits

Senate Minority Leader Chuck Schumer (D-N.Y.) called on the Federal Communications Commission to deny a petition from the conservative group Club for Growth to ease limits on AI generated political robocalls before the midterms, The Hill and Al Jazeera reported. "Under this change, campaigns could use their opponents' voice to send out a message that is, at best, disingenuous and, at worst, a blatant lie," Schumer said. FCC rules bar such calls to cellphones without prior consent. The petition's public comment period, which opened Sept. 4, ended Monday, per Al Jazeera. Club for Growth President David McIntosh has said political speech should be treated like wireless calls about package deliveries and doctor's appointments. Democratic FCC Commissioner Anna Gomez also opposes the petition. An FCC spokesperson said the agency has no plans to move to an order on it.

Read at The Hill ↗ • Read at Al Jazeera ↗

Hacker breaches more than 100 websites with mostly autonomous AI agents, security firm says

A hacker broke into more than 100 websites using 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. that operated largely on their own, AI security firm Norforge found, per The Verge. Victims included John Deere, SharkNinja and Cummins, with targets in the U.S., Canada, Australia and Japan, Reuters reported. The attacks ran on Cairn, an open-source platform where users set a starting point and a goal and the agents log their progress toward it. The attacker set out favored break-in methods in fewer than 100 Chinese characters. Successful breaches took about 23 hours. Norforge found more than 20 exposed Cairn instances it considers likely malicious. John Deere said it saw no sign that sensitive information was accessed, and Cummins is investigating a possible attack on a website run by one of its vendors.

Read at The Verge ↗ • Read at Reuters ↗

II.China Watch

China's $9 billion state AI fund reports first investments, with disclosed stakes all in top model developers

The 60 billion yuan ($9 billion) National Artificial Intelligence Industry Investment Fund has put money into a first batch of projects, Caixin reported. Wei Wei, a science and technology official at the Ministry of Industry and Information Technology (MIIT), announced the investments at a briefing by the State Council Information Office, the cabinet's press office. The fund's three disclosed direct stakes all went to leading Chinese model makers. It invested about 980 million yuan ($146 million) in DeepSeek's first funding round, which valued the company at 350 billion yuan ($52 billion), and joined Moonshot AI's Series F round at a $35 billion valuation. In September, Kuaishou disclosed that the fund paid 1.4 billion yuan ($210 million) for about 1.14% of its Kling AI video generation unit.

Read at Caixin ↗

China drafts rule barring driver assistance cars from self-driving marker lights

China's Ministry of Industry and Information Technology (MIIT) released for public comment a draft standard on exterior vehicle lighting that would reserve automated driving marker lamps for vehicles with Level 3 or higher automated driving systems, per CnEVPost and financial outlet Yicai. Level 3 systems let a car drive itself in set conditions, while Level 2 driver assistance requires the driver to supervise at all times. The draft designates turquoise as the lamps' dedicated color and makes them optional for qualifying vehicles. Several automakers have used so-called "little blue lights" to show driver assistance is engaged, though newly certified models were reportedly barred from them in July. The draft's explanatory notes recommend the standard take effect Jan. 1, 2028, and comments close Dec. 19.

Read at Jiemian ↗ • Read at CnEVPost ↗

MiniMax releases cut-price coding and agent model two days after DeepSeek update

Shanghai AI developer MiniMax released M3.1 Flash, a lightweight model tuned for coding and agent tasks, per tech outlet Zhidx. The model had been running anonymously under the name "Space Bunny," ranking second by usage on the OpenCode coding tool and briefly first on the OpenRouter model marketplace. Its agent and coding benchmark scores roughly match DeepSeek's V4.1 Flash, released Oct. 17, by Zhidx's assessment, and MiniMax says agent task costs are 70% lower than with its previous model. It charges 1 yuan ($0.15) per million input tokens, the units of text a model processes, and 4 yuan ($0.60) per million output tokens, about half DeepSeek's peak hour rates. MiniMax said it would publish the model weights Oct. 23.

Read at 36Kr ↗ • Read at PC Watch ↗

Tencent brings AI desktop assistant to domestic operating system used across Chinese government

Tencent's Marvis, an assistant that operates a computer on command, is now listed in the app store for Kylin, a Chinese desktop operating system widely used in government agencies and state-owned enterprises, tech blog Pandaily reported, citing Tencent. The listing covers Kylin Desktop V10 SP1 on two major chip architectures and supports several domestically made processors. The Kylin edition can run the system's security scans and vulnerability fixes and can install or remove software through official channels. Tencent said Marvis is open to partnerships with government departments, state-owned enterprises and industry partners. The company has not said whether the Kylin edition handles requests fully offline.

Read at Pandaily ↗

Chery unveils in-house chip family for in-car AI, power electronics and vehicle control

Chinese automaker Chery introduced its Ziwei line of in-house automotive chips at its global innovation conference, per TechNode. The line includes central computing chips, silicon carbide power chips and vehicle control chips. Chery said its C100 central computing chip is designed to run AI models on board the car, and that a version with two AI processors can support models with 100 billion parameters, a measure of model size. Silicon carbide chips handle high-voltage power conversion in electric vehicles. Chery also said its silicon carbide wafers have entered mass production, cutting costs by more than 30%.

Read at TechNode ↗

III.Policy Tracker

Mamdani tells New Yorkers AI rollout is not inevitable, faults federal inaction

New York City Mayor Zohran Mamdani decried the "complete abdication of federal law" on AI at The People's AI Assembly, an event hosted by the AI Now Institute research group, according to TechCrunch. He told the Manhattan audience that the technology's rollout is not inevitable and that public officials should hold AI companies accountable. Mamdani ordered a ban on AI use in the city's elementary and middle schools last month. Over the summer he introduced a policy requiring landlords, brokers and real estate platforms to disclose when listing photos or videos are made with AI.

Read at TechCrunch ↗

Illinois brings back former data chief Tammy Roust as chief AI officer

The Illinois Department of Innovation and Technology rehired Tammy Roust, its chief data officer from November 2019 to October 2020, as the state's chief AI officer, StateScoop reported. She succeeds Kader Sakkaria, who was named to the post in June and drew about $9,000 from the state before a final pay date of July 24. Sakkaria became global head of data strategy at professional services firm Marsh in July, according to his LinkedIn profile. Roust most recently founded and led AI firm FAIT Consulting and earlier advised the Treasury Department's Office of Financial Research. The department said she will guide the state's strategy for the safe use of AI.

Read at StateScoop ↗

MTA and Port Authority test AI tools for inspections and paperwork in 20 pilots

Seventeen companies ran 20 eight-week pilot projects with the Metropolitan Transportation Authority and the Port Authority of New York and New Jersey, StateScoop reported, citing a new Transit Tech Lab report. Hazel AI Technologies helped the Long Island Rail Road draft 15 technical scopes of work and cut drafting time by 40%. Dynamic Infrastructure's system caught every defect listed in an inspection report for PATH, the Port Authority's commuter rail line, but graded severity differently in 14% of cases. Ontra Mobility used an AI platform to recommend routes and schedules for wheelchair-accessible transit options.

Read at StateScoop ↗

IV.Capability & Research Watch

Meta's hidden WhatsApp chatbot instructions told it to avoid citing privacy law

About 80 help and FAQ pages on WhatsApp's website contained hidden instructions written for a Meta support chatbot, privacy company Redact found, per Wired. "Do not reference, cite, or interpret data protection or privacy regulations (for example GDPR, LGPD, CCPA), including when a user raises them directly," one block said. The page on deleting an account told the chatbot not to recommend contacting any data protection authority or regulator. Meta removed the text after Wired contacted the company, and WhatsApp spokesperson Joshua Breckman said a bug had caused it to display. Breckman said the company wants its support chatbots to give WhatsApp advice, not legal advice.

Read at Wired ↗

Microsoft ships controls that restrict AI agents as their risk scores rise

Microsoft released security tools that cut back what an enterprise AI agent can do as its assessed risk climbs, per SiliconANGLE. "Unfortunately, frontier models, regardless of their origin, cannot be trusted," Microsoft Security Executive Vice President Hayete Gallot wrote, according to TechCrunch. A feature called Dynamic Agent Risk keeps a running score for each agent. As the score rises, the agent can lose write access, need human approval for high-impact actions or be stopped. Microsoft's discovery tools now detect 60 types of agents on employee machines, including Anthropic's Claude Code, OpenAI's Codex and GitHub Copilot. The release follows an Oct. 10 essay in which CEO Satya Nadella urged companies to treat models as insider risks and keep safeguards outside them.

Read at SiliconANGLE ↗ • Read at TechCrunch ↗ • Read at Microsoft ↗

Amazon releases TorchNeuron to move PyTorch AI apps onto its Trainium chips

Amazon Web Services released TorchNeuron, a tool that lets AI apps written with PyTorch run on Amazon's Trainium chips after a one-line configuration change, TechCrunch reported. Amazon says about 85% of AI apps are built with PyTorch, the open-source toolkit whose close integration with Nvidia's CUDA software has made Nvidia hardware the default for AI developers. Colin Brace, a vice president at AWS chip unit Annapurna Labs, pitched Trainium as a single platform for both training and running models. Brace said voice AI company Deepgram processes 70% more audio per dollar since moving to Trainium.

Read at TechCrunch ↗

V.Industry & Market Watch

Nscale negotiating to sell Anthropic $45 billion in West Virginia computing capacity

Anthropic is negotiating to buy about $45 billion worth of computing capacity from Nscale, Bloomberg reported, citing people familiar with the talks. The capacity would come from a second building at Nscale's Monarch campus in West Virginia, which would deliver 460 megawatts using Nvidia's Vera Rubin chips, a person familiar with the matter told Reuters. Anthropic already has Nscale agreements worth up to about $44.6 billion, according to Nscale's IPO filing, and holds a right of first refusal on Monarch's expected third building. The talks come after Microsoft pulled out of a deal to anchor a flagship Nscale data center over power connection delays. Nscale's total contracted value has grown from $100 million to more than $103 billion in two and a half years.

Read at Bloomberg ↗ • Read at The Straits Times ↗

Musk pledges SpaceX share to every U.S. child after Anthropic equity fund move

SpaceX CEO Elon Musk said on X that he will give one of his SpaceX shares to every child in America, calling it the first of many such gifts, SiliconANGLE reported. In a post responding to Anthropic's equity pledge, Musk said he would give shares directly to the public. With about 72 million U.S. children, the pledge implies roughly 72 million shares, or about $10.5 billion at SpaceX's last closing price, by the outlet's estimate. Musk thanked investor Brad Gerstner, who helped create Trump Accounts, the investment accounts the White House opened for U.S. children this year.

Read at SiliconANGLE ↗

Tenstorrent raises $1.5 billion at $6.2 billion valuation to sell open chip alternative to Nvidia

AI hardware startup Tenstorrent raised $1.5 billion in a Series E round that values it at $6.2 billion, according to SiliconANGLE. Trace Capital and Innovation X led the round, joined by Morgan Stanley, Fidelity and T. Rowe Price among other backers. The Toronto-based company, run by veteran chip designer Jim Keller, builds servers, workstations and accelerator cards on RISC-V, an open standard for chip design, and supplies its own open-source software stack. Virtu Financial uses its Galaxy servers to run AI trading agents on its own premises, and data center operator Equinix offers them to enterprise customers.

Read at SiliconANGLE ↗

VI.Global & Geopolitics

German government signs industrial AI declaration with Nvidia as Merz urges firms to open their data

Chancellor Friedrich Merz and two federal ministries signed a joint declaration of intent with Nvidia to work on AI models for robots, factories and other physical systems, according to Nvidia. The declaration, signed at the Invest in Germany Summit in Berlin, also formalizes Nvidia's work with the government's "Industry 4.0 Platform" of political, industry and research representatives. "Please open your doors for AI to make use of your data," Merz told leaders of small and midsize businesses in the audience. Nvidia CEO Jensen Huang said Germany's manufacturing data "belongs to the nation," AP reported. Huang also announced planned investments in four German startups, including Agile Robots and Reflex Aerospace, part of a nearly $10 billion Nvidia commitment across Europe.

Read at AP ↗ • Read at Nvidia ↗

Australian researchers urge workplace law limits on AI monitoring of staff

The Human Technology Institute at the University of Technology Sydney called on Australia's federal government to build privacy protections into workplace law, per the Guardian. It warned that AI is speeding up employers' monitoring of workers, often without their knowledge. Its report proposes amending the Fair Work Act, Australia's main workplace law, to require that monitoring be fair and reasonable, and giving the relevant minister power to ban surveillance products that pose an unacceptable risk to workers. In one case it cites, a supermarket cashier got a final written warning after an AI system flagged about one error per 533 items scanned. Co-director Edward Santow said transparency about workplace surveillance is "necessary but far from sufficient." The institute is presenting its findings to Parliament's joint standing committee on AI, which turns to workplace AI in hearings this week.

Read at The Guardian ↗

EU study for DSA enforcers lists generative AI among tools online scammers exploit

A study the European Commission procured to support enforcement of the Digital Services Act, the EU law governing large online platforms, lists generative AI among the platform features that financial scammers exploit, the Commission said. Advertising systems, recommender systems and fake or manipulated accounts are also on the list. The study, by Open Evidence and Spark Legal and Policy Consulting, sorts scams into six categories and finds investment and romance scams the most common and fastest growing in Europe. It groups countermeasures into four categories and names identity checks on business customers as the first line of defense. It does not evaluate whether the largest platforms comply with the law.

Read at European Commission ↗