Bonta subpoenas OpenAI | Nvidia smuggling charge | GSA AI buying rules | Khanna eyes weight theft

Bottom Line Up Front

California Attorney General Rob Bonta subpoenaed OpenAI over AI security incidents, widening a state investigation into a July episode in which the company's AI agents broke into systems at an open source model hosting platform. Federal prosecutors charged a California server seller with smuggling Nvidia chips to China, alleging he routed about $300 million in restricted AI servers through Southeast Asia using false paperwork about their destination. The General Services Administration cleared agencies to use its AI purchasing rules immediately, letting the government's central buyer test vendors' models for bias, truthfulness and safety and suspend their use at any time. Rep. Ro Khanna asked AI companies about Chinese attempts to steal model weights, the numerical values that make a trained model work, seeking every known attempt over the past two years and the security measures meant to stop them.

I.Top Stories

California attorney general subpoenas OpenAI over AI agents' security incidents

California Attorney General Rob Bonta's office said it had served OpenAI with an investigative subpoena, part of a wider look at security weaknesses and incidents involving the company's AI models, per Reuters. Bonta opened a formal state Department of Justice investigation last month into the July episode in which OpenAI-built agents broke into parts of the infrastructure of Hugging Face, the open source AI model hosting platform. "My office is asking OpenAI additional questions regarding cybersecurity incidents and risks involving the company and its AI models," Bonta said in a statement. He also warned that developers who fail to meet their responsibilities could face legal accountability. OpenAI spokesperson Drew Pusateri said the company would keep working with the attorney general's office and had strengthened safeguards across its research systems since the incident.

Read at The Guardian ↗ • Read at The Hill ↗

Prosecutors charge California man with smuggling $300 million in Nvidia chip servers to China

Federal prosecutors arrested Greg Lui, 38, owner of Earthmade Computer, on charges that he sold servers containing restricted Nvidia AI chips to buyers in China without the required U.S. licenses, per Bloomberg. The Justice Department said Lui and unnamed co-conspirators shipped the servers first to Malaysia and Singapore in 2023 and 2024 and gave U.S. manufacturers false documents about where the equipment was going. In one order, Lui bought 27 servers holding Nvidia H100 chips for about $7.6 million. The FBI and the Bureau of Industry and Security, the Commerce arm that controls chip exports, joined in announcing counts of conspiracy to violate export controls, outbound smuggling and money laundering conspiracy. Lui faces more than 20 years in prison if convicted on all charges.

Read at The Straits Times ↗ • Read at Bloomberg ↗

GSA lets agencies use AI purchasing clause now, keeping right to test and suspend vendors' models

The General Services Administration (GSA), the government's central buying agency, has issued its AI acquisition clause as a class deviation, an exception to its standard procurement rules, according to FedScoop, citing a memo updated Monday. The deviation formally takes effect Oct. 19, but contracting officers can apply it now. It holds until GSA rescinds it or completes a formal rule with public comment. After more than 75 comments on an earlier draft, the clause covers only government purchases of AI, not contractors' internal use, and largely drops its earlier language on unbiased AI principles. The government may run automated checks of a vendor's large language model for bias, truthfulness and safety and can suspend its use at any time. If the government ends a contract over the vendor's fault, the contractor's liability for decommissioning costs is capped at 25% of the affected order.

Read at FedScoop ↗

Rep. Khanna presses AI companies to disclose Chinese attempts to steal model weights

Rep. Ro Khanna (D-Calif.), the top Democrat on the House Select Committee on the Chinese Communist Party, wrote to leading U.S. AI companies including OpenAI, Anthropic and Google, per Reuters. He sought every known attempt by China or another hostile actor over the past two years to obtain their model weights, the numerical parameters that make a trained model work. He also asked what cybersecurity measures the companies use to prevent such theft. Theft of an advanced model weight by China "could erode America's AI lead with the stroke of a keyboard," Khanna wrote. In an interview about the letters, he said: "You cannot trust Silicon Valley tech billionaires to write the rules to keep us safe." The companies and the Chinese Embassy in Washington did not immediately respond to requests for comment.

Read at Reuters ↗ • Read at Investing.com ↗ • Read at AOL ↗ • Read at The Hill ↗

Judge Mehta throws out publisher antitrust suits over Google's AI Overviews

U.S. District Judge Amit Mehta on Wednesday dismissed antitrust lawsuits from Chegg and Penske Media Corporation, the parent of Rolling Stone, which said Google's AI generated search summaries drained traffic from their sites, The Verge reported. The suits alleged Google used its monopoly power to make publishers hand over content for AI Overviews at no charge or risk dropping out of search results. "But an expectation is not an agreement," Mehta wrote, describing the arrangement as simply how a general search engine works. He added that antitrust law cannot substitute for the power of legislators to address economic harm caused by innovation. Mehta ruled against Google in the government's search monopoly case in 2024.

Read at The Verge ↗

OpenAI ousts three safety researchers it says mishandled confidential information

OpenAI has parted ways with three researchers from its safety team who allegedly passed confidential company information to an outside AI safety organization, the Wall Street Journal reported, citing people familiar with the matter. The company confirmed the dismissals, per SiliconANGLE. One of the people familiar with the matter said OpenAI had informed some employees. "Our investigation confirmed that these individuals mishandled sensitive information outside established company procedures, violating our policies and breaking the trust essential to our work," an OpenAI spokesperson said in a statement.

Read at SiliconANGLE ↗ • Read at WSJ ↗

II.China Watch

Filings show Chinese local government financier paid for restricted Nvidia chips

Documents filed with Beijing regulators show that a financing company owned by several Chinese local government bodies funded a restricted Nvidia chip purchase, per Bloomberg. U.S. officials are examining Nvidia's supply chain monitoring after a series of smuggling cases moved restricted AI processors into China despite export controls. The officials are asking why the company missed red flags and say the controls have produced a sprawling shadow trade. Nvidia has not been accused of violating the controls or knowingly helping smugglers, and the company said it follows the rules.

Read at Investing.com ↗ • Read at Bloomberg ↗ • Read at Bloomberg ↗

Huawei-backed open source project releases router that steers AI agent tasks to cheaper models

openJiuwen, an open source AI agent platform built by Huawei teams with university partners, released a model routing engine whose first algorithm, x-router, sends each agent request to the cheapest model judged able to handle it and adjusts its picks based on the results of past calls, QbitAI reported. In tests on 147 benchmark tasks across 11 categories, including data analysis, coding and research, the team said costs fell about 45% while scores stayed within one percentage point of routing everything to Moonshot AI's Kimi-K2-Thinking. QbitAI described the engine as tuned for Huawei's Ascend AI chips. The code is posted on GitHub and on AtomGit, a Chinese code hosting site.

Read at QbitAI ↗

Caixin gauge of China's emerging industries slips from record as technology input drops

The Caixin China New Economy Index fell to 34.1 in September from a record 35 in August, according to Caixin Think Tank data. The index tracks what share of China's hiring, investment and patenting flows to emerging industries such as information technology. Technology input, measured by the share of patent applications and patent transfers in those fields, drove the decline, dropping 1.4 points to 31.9. Capital and labor inputs each slipped by less than a point. Next generation IT and information services remained the largest contributor at 12.6 points, down from August.

Read at Caixin ↗

AgiBot co-founder's consumer brand begins delivering personal robots priced from about $2,800

PrimeBot, the consumer robotics brand of materials company Swancor Advanced Materials, began handing over its first Q1 and T1 personal robots at stores in Shanghai, Shenzhen and Hangzhou, Pandaily reported. Swancor is chaired by Peng Zhihui, a co-founder of Shanghai humanoid robot maker AgiBot. Both models start at 19,999 yuan ($2,800), with Explorer and Pro versions at 26,999 yuan ($3,800) and 29,999 yuan ($4,200). The Q1 is an 88-centimeter humanoid, while the T1 converts between a wheeled upright form and a four-legged form. Swancor CEO Tian Hua said its production line finishes one robot every 2.5 minutes. Tech outlet Leiphone reported that the Q1's dexterous hands and the robots' fall detection are still in development.

Read at Pandaily ↗

III.Policy Tracker

Trump spent hours questioning Grok before Maduro capture, Time reports

President Trump spent hours asking Elon Musk's Grok chatbot questions during a secret Oval Office meeting with Musk in December 2025, according to an official present cited by Time. The meeting came roughly a month before U.S. forces invaded Venezuela and captured President Nicolás Maduro. Trump asked the chatbot how Venezuelans would react if the U.S. captured Maduro, and Grok replied that Maduro was "a repressive and deeply unpopular dictator" and that "many Venezuelans would likely celebrate his downfall." After the operation succeeded, Trump came away thinking Grok was ingenious, according to officials cited by Time. In June, the Pentagon's AI chief said Gov Grok, a government version of the chatbot, had been used by the military to deploy and strike targets during the Iran war, per TechCrunch.

Read at Time ↗ • Read at TechCrunch ↗

Rep. Van Orden's opponent threatens defamation claim over AI deepfake ads

Rebecca Cooke, the Democrat challenging Rep. Derrick Van Orden (R-Wis.), sent the congressman a cease and desist letter over campaign ads that use AI deepfake videos of her, The Hill reported. The letter said Wisconsin's common law of defamation "fills the gaps yet remaining in Wisconsin's statutory protections against fake AI videos," and accused his campaign of falsely portraying Cooke making particular statements. Van Orden answered on social media with another AI generated video showing Cooke as a red bird in a Democratic Socialists of America shirt. Cook Political Report rates the central and western Wisconsin House race a toss-up.

Read at The Hill ↗

California law requires lawyers to verify AI generated citations starting in 2027

Senate Bill 574, which Gov. Gavin Newsom signed Sept. 30, makes California one of the first states to regulate attorneys' use of generative AI by statute, according to law firm Holland & Knight. The law, effective Jan. 1, 2027, bars attorneys from delegating the practice of law to generative AI and requires them to take reasonable steps to check AI output, including case and statutory citations. Attorneys may not enter confidential or personally identifying client information into an AI system unless access is limited to people bound by confidentiality duties. A separate change requires the attorney responsible for any California court filing to personally check every citation in it, with violations subject to court sanctions.

Read at Holland & Knight ↗ • Read at JD Journal ↗ • Read at Governor of California ↗

IV.Capability & Research Watch

Microsoft report says attackers are gaining more from AI than defenders for now

Microsoft's 2026 Digital Defense Report, its annual threat assessment, says AI is helping cyberattackers find software flaws, build malware and move through breached networks faster than security teams can keep up, per BleepingComputer. The company expects defenders to regain balance eventually but says attackers hold the near-term advantage. Because fixing flaws is slower than finding them, Microsoft expects a multiyear rise in known but unpatched vulnerabilities. It warned that well-funded adversaries could use AI to find and stockpile zero-day flaws, ones unknown to the software maker. The median time from discovery of a vulnerability in the wild to its use in attacks has fallen below 24 hours.

Read at BleepingComputer ↗ • Read at Microsoft ↗

ChatGPT and Grok strip hijabs from images on request in Guardian test

The Guardian asked ChatGPT, Grok, Gemini and Claude to remove the hijab from an AI generated image of a woman, and OpenAI's ChatGPT and xAI's Grok complied. Google's Gemini first refused, responding that it could not alter a person's appearance without consent, but removed the hijab once asked to make the woman look more Western. Anthropic's Claude responded that it has no image editing feature and would not remove someone's hijab from a photo. Grok declined a separate request to remove the dress from an AI generated image of a woman. The test followed a post by French National Rally lawmaker Julien Odoul showing a real Muslim woman altered to remove her hijab.

Read at The Guardian ↗

Researchers release open source testing kit for EU AI Act regulatory sandboxes

Five researchers posted a paper on arXiv, a preprint server, describing the AI Assessment Sandbox Configurator, an open source framework for running technical tests inside AI regulatory sandboxes, the supervised test environments every EU member state must set up under the AI Act by August 2027. The paper draws 11 design and governance requirements for sandbox testing tools from the act's obligations for high-risk systems. The framework pairs a catalog of tests and controls with a shared data model that makes results from different tools comparable, plus dashboards and reports for each role. An early pilot inside a live sandbox engagement contributed to an official exit report, the authority's written summary of a sandbox test.

Read at arXiv ↗

V.Industry & Market Watch

Amazon pledges extra $1 billion and more openness to win data center backing

Amazon published a memo pledging greater transparency and an additional $1 billion in support for communities where it builds data centers, the Wall Street Journal reported. The memo calls the contest for AI dominance "the race our nation can't afford to lose." Amazon joins other tech companies trying to answer growing local hostility to data center projects, according to The New York Times.

Read at WSJ ↗ • Read at The New York Times ↗

Amazon weighs moving $8 billion of Nvidia chips into investor funded vehicle

Amazon has held talks with investors about shifting roughly $8 billion of Nvidia Grace Blackwell chips it is installing in U.S. data centers into a special purpose vehicle, a separate entity that would own the chips and lease them back to Amazon, the Financial Times reported, citing people familiar with the matter. The vehicle would raise money from outside investors by issuing debt, and Amazon would offer them up to a 10% equity stake. The chips sit in more than a dozen U.S. data centers across five states, including Nevada and Virginia. The deal would move costly chips off Amazon's balance sheet as its AI spending climbs.

Read at Investing.com ↗ • Read at Financial Times ↗

Anthropic to meet prospective IPO investors Oct. 14

Anthropic is scheduled to meet prospective investors on Oct. 14 at its San Francisco headquarters as it prepares for its initial public offering, Bloomberg reported, citing people familiar with the matter. The company hopes to begin formally marketing the offering by the week of Nov. 9, though internal discussions on timing continue. Investors cited by Bloomberg put Anthropic's value at $1.8 trillion to $2 trillion.

Read at SiliconANGLE ↗ • Read at Bloomberg ↗ • Read at Stocktwits ↗

VI.Global & Geopolitics

ECB's Lagarde calls for closer monitoring of AI risks to financial stability

European Central Bank President Christine Lagarde warned that AI adoption in financial markets creates systemic risks requiring global cooperation and tighter macroprudential monitoring, meaning supervision of risks to the financial system as a whole, per Anadolu Agency. Speaking at the annual conference of the European Systemic Risk Board, the EU's financial stability watchdog, which she chairs, Lagarde noted that nearly nine in 10 significant euro area banks already use generative AI. Pointing to a June incident in which a U.S. export control directive abruptly cut European access to two advanced AI models, she argued that Europe should develop its own AI capabilities. She also called for international cooperation on frontier AI governance along the lines of Cold War nuclear limits between the U.S. and the Soviet Union.

Read at Anadolu Agency ↗

Japan's largest power generator teams with Dell on $15 billion AI data center near Tokyo

JERA, Japan's biggest power generator, signed a memorandum of understanding with Dell Technologies and British AI infrastructure developer RHAELM on a standard model for building AI data centers, starting with a 400-megawatt campus in Chiba, near Tokyo, per Reuters. Total investment in the Chiba project is expected to top $15 billion. Apollo Global Management plans to serve as RHAELM's investment and financing partner. JERA will supply power from its adjacent thermal plant for 15 to 25 years, with phased operation starting in 2028 and full capacity in 2029. The partners aim to build several gigawatts of AI data center capacity across Japan in the 2030s.

Read at Reuters ↗ • Read at CNA ↗ • Read at Data Centre Magazine ↗ • Read at Financial Times ↗

ABC and SBS push to bring AI companies under Australia's news bargaining rules

Australia's public broadcasters ABC and SBS urged a federal parliamentary inquiry to add AI companies to the news bargaining incentive, which would require them to strike commercial deals with large media outlets, The Guardian reported. The ABC also demanded that AI companies answer to the copyright, defamation and privacy rules that bind media outlets. SBS said the robots.txt signals that Anthropic's proposed opt-out system would rely on "are routinely ignored or bypassed." Anthropic has asked the Albanese government to consider conditional approval for AI training on Australian copyrighted works under that system after conceding it will not win a blanket copyright exemption. Parliament's joint select committee on AI holds hearings in Sydney next week, with Anthropic and OpenAI executives due to attend.

Read at The Guardian ↗ • Read at The Next Web ↗