Silicon Valley Daily
A Defense Department official told the BBC on Monday that the Pentagon ‘has ceased the use of Anthropic products’; sources said Claude was still used as recently as last week, including in operations against Iran. In New York, Anthropic, OpenAI, Google and Meta testified under oath before the City Council, and SpaceXAI did not appear despite a subpoena, according to Speaker Julie Menin. OpenAI made two product announcements: invisible text watermarks for ChatGPT and Codex in the EU, and visual ads during image generation in the US. In chips, two former Groq engineers sued over Nvidia’s $20 billion licensing deal, the Financial Times reported that Tencent will rent about 100,000 AI chips from Oracle for roughly $7 billion, and Huawei and Qualcomm signed a patent cross-license. 404 Media reported that Meta fixed a virtual-machine escape flaw in Muse shortly before launch, and Denmark disclosed that data on 8.8 million people was taken from its population register.
Pentagon tells BBC it has stopped using Anthropic’s AI tools
A Defense Department official told the BBC in a statement on Monday that the Pentagon ‘has ceased the use of Anthropic products’. Defense Secretary Pete Hegseth designated Anthropic a supply chain risk in February and said use would end by late August. People familiar with the matter told the BBC that Claude was still in use as recently as last week for research, analysis and intelligence gathering and in military operations against Iran, embedded in the Palantir-operated Maven Smart System. Anthropic declined to comment; it has called the designation ‘unprecedented and unlawful’ and is suing the administration. Two people told the BBC that OpenAI tools have been adopted more widely in some military departments in recent months.
AI labs testify under oath at New York City Council hearing
Representatives of Anthropic, OpenAI, Google and Meta testified under oath on Monday at a New York City Council hearing on AI risks, CNBC and Bloomberg report. Former Anthropic researcher Jacob Coxon told the council that ‘it is more likely than not that humanity loses control to these AIs’ on the current path. Former Google DeepMind researcher Alex Turner said the United States is ‘racing to build and grow our own adversary’ in misaligned AI. Council Speaker Julie Menin called the companies’ answers on legal liability ‘troubling at best’ and said SpaceXAI stayed away in ‘direct violation of the subpoena’. According to Menin, Anthropic, OpenAI and Google agreed to appear only after the council threatened subpoenas.
OpenAI to watermark ChatGPT and Codex text in the EU
OpenAI said it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks to comply with the EU AI Act. API customers worldwide can opt in for select models starting today; watermarking stays off by default in the API. Access to the detector is limited at first to approved researchers and expert organizations. The Decoder, citing OpenAI’s figures, reports that the textGrain detector caught about 95 percent of 400-token passages in one test, and 17 percent once a quarter of the words were replaced.
OpenAI to test visual ads during ChatGPT image generation
OpenAI announced a visual ad format for ChatGPT on Monday and said testing will begin in the US later this month with an initial group of advertisers. The ads will first appear during image generation, will be labeled and will stay separate from the generated image; OpenAI says advertising does not influence ChatGPT’s answers. The company also added measurement integrations with Hightouch, Tealium and LiveRamp. In the announcement, OpenAI says ChatGPT reaches 1.2 billion people each week. The Verge notes that ChatGPT ads began in February and until now showed only a company’s name, logo and link beneath the chat.
404 Media: Meta rushed to fix Muse VM escape flaws before launch
404 Media reports that Meta engineers found several security vulnerabilities in its Muse AI agent in the weeks before launch. According to a Meta source and internal documents seen by the outlet, at least one flaw could have let an ordinary Muse user break out of the product’s virtual machine and reach sensitive internal Meta databases. An internal post described a ‘mad dash’ to fix ‘a sudden spike in reported KVM escapes’, and the issue was raised to Mark Zuckerberg. The flaws were found and fixed before release. A Meta spokesperson told 404 Media the company is proud of the work done to make Muse ‘safe, secure and private’.
Former Groq engineers sue over Nvidia’s $20 billion deal
Former Groq engineers Joshua Rubin and Benjamin Serebrin filed suit in the Delaware Court of Chancery on Oct. 2, CNBC reports. They allege that Groq’s board ‘sold the company to Nvidia without the stockholder vote Delaware law requires’ and that stockholders received a ‘lowball’ price. The complaint says Nvidia allocated $17 billion to a license labeled non-exclusive and $3 billion in restricted stock units for Groq employees who moved over, and that about 150 to 200 engineers joined Nvidia. Groq called the lawsuit ‘meritless’ and said the licensing agreement delivered ‘exceptional value’.
FT: Tencent to rent 100,000 AI chips from Oracle for about $7 billion
Tom’s Hardware, citing the Financial Times, reports that Tencent has agreed a five-year lease covering about 100,000 advanced AI chips at Oracle data centers in Southeast Asia. The deal is estimated at about $7 billion, with roughly 30 percent paid upfront. The chips are unavailable in China, and neither company has named them or commented. The FT says such leases are legal under current US rules. Tom’s Hardware calculates the price at about $1.60 per chip per hour.
Huawei and Qualcomm sign multi-year patent cross-license
Huawei and Qualcomm announced a multi-year patent license agreement on Monday. It includes cross licenses covering 5G, compute, AI and networking. Qualcomm will also buy certain Huawei US patents in compute, AI, networking and other fields, a transaction that closes after regulatory approvals. Financial terms were not disclosed in Huawei’s announcement.
Report: Chinese chipmakers imported 343 ASML immersion DUV scanners
A report from the Washington-based Centre for Technology & Statecraft estimates that Chinese logic and memory chipmakers imported 343 ASML immersion DUV lithography scanners between 2012 and early 2026, Tom’s Hardware reports. CTS says 270 of them are Twinscan NXT:1980i machines that can be used for 7nm-class production through multipatterning. It estimates that Chinese fabs bought about 90 of that model in 2024 and 89 in 2025. The group calls for banning exports of all immersion DUV tools to China, in line with the MATCH Act introduced in April. Tom’s Hardware says some of the report’s estimates look questionable.
Altman gives Vanity Fair and Politico interviews; publicist interrupts suicide question
Vanity Fair published a video interview with Sam Altman covering President Trump, the midterms, extinction risk and self-policing versus regulation. The Verge reports that when editor Mark Guiducci raised a ChatGPT user’s suicide, an OpenAI publicist said she would ‘like to move on’ because time was short. Guiducci continued, and Altman called such cases ‘some of the hardest questions that we face’. OpenAI spokesperson Drew Pusateri said the interview was running over and that more time was later scheduled. In Politico’s Decoded podcast, released Monday, Altman described OpenAI as ‘pragmatic centrists’ on regulation and repeated that the world should accept ‘some bad things happening’ for AI’s benefits.
Denmark says population registry breach exposed 8.8 million records
Denmark’s government said unauthorized access to the Central Person Register exposed information on about 8.8 million registered people, including residents, people who moved abroad and the deceased. The register holds about 11 million records. According to the ministry’s statement, the attackers abused a Danish company’s lawful access to look up names, addresses and CPR numbers. The access happened in September and was discovered on Oct. 2; the company’s access has been blocked and police are investigating. Minister Christina Egelund called it an extremely serious incident. TechCrunch’s report puts the figure at about 8 million.
Utah lets Nolla Health’s AI issue acne prescriptions in pilot
Nolla Health said on Monday that Utah’s Office of Artificial Intelligence Policy has authorized a pilot in which its Nolla Derm app issues initial acne prescriptions to adult Utah residents. Bloomberg describes it as the first US pilot of AI diagnosing and prescribing without direct human oversight. The company’s plan is staged: for the first 100 patients, two licensed physicians approve every prescription; for up to 500 patients, a physician reviews every case afterwards; later, physicians sample at least 10 percent each month. Moving between stages requires 95 percent agreement with physicians, zero serious adverse events and written state approval. The company says Utah adults can sign up now, priced from $4.99 a month.
Mortgage software firm Valon raises $150 million at $2.3 billion valuation
Valon announced a $150 million Series D at a $2.3 billion valuation, which the company says doubles its previous valuation. Ribbit Capital joined as a new investor, with Andreessen Horowitz participating again. Valon sells ValonOS, an operating system for mortgage servicers, and says it partners with servicers responsible for one in six US mortgages. Forbes reported the round first.
German robotics startup RobCo passes $1 billion valuation in share sale
RobCo said it has surpassed a $1 billion valuation in a transaction that brings in new investment and lets long-standing employees sell shares. The company says its valuation doubled in nine months; Sifted notes it raised $100 million in January. Existing investors including Sequoia and Lightspeed took part alongside new investors such as Cherry Ventures. RobCo plans to commercially launch Alfie, a robot for industrial work, in Munich on March 4, 2027.
Google pauses open-source bug bounty over automated submissions
Google has stopped accepting product vulnerability submissions to its Open Source Software Vulnerability Rewards Program as of Oct. 1. The company said the pause is ‘due to a significant rise in automated submissions, the vast majority of which are not valid’, and committed to an update in the first quarter of 2027. Supply chain reports and outstanding reports are not affected, BleepingComputer reports. The program launched in August 2022 with rewards from $100 to $31,337.
New York Magazine interviews researchers who left AI labs
New York Magazine published exit interviews with researchers who left OpenAI, Anthropic and DeepMind, among them Jacob Coxon, Daniel Kokotajlo and Alex Turner, on AGI risks, planned IPOs and oversight. Coverage of David Robinson, who resigned from OpenAI’s safety team last week, continued. Tom’s Hardware notes that his Atlantic essay cites the July Hugging Face hack and a failed ‘kill switch’, and urges the industry to learn from fields such as nuclear engineering and aviation. Business Insider reports he left the same week three other safety researchers were fired, and quotes former colleague Joshua Achiam calling him ‘sober and thoughtful’.
Norway proposes temporary ban on camera glasses in some public places
Norway’s government said on Monday it will send parliament a law temporarily banning the use of AI glasses in places such as parks, beaches, schools and kindergartens, Ars Technica reports. An expert group will propose permanent rules. Digitalization minister Torgeir Micaelsen said people risk ‘being photographed, filmed, or recorded without their knowledge’. Business Insider notes that sales of Meta’s glasses and private use are not banned. Equinor has already barred the devices from its offices and offshore facilities.
Amazon drops data center NDAs, pledges over $1 billion to host communities
Amazon said it will no longer require nondisclosure agreements with government agencies involved in its data center projects, Tom’s Hardware reports. Under a new Data Center Commitment program it will invest more than $1 billion over five years in host communities and fund efficiency upgrades for 30,000 homes. It will also disclose energy and water use annually. In a separate post, AWS chief executive Matt Garman wrote that if the roughly 100 data center moratoriums under consideration pass, ‘the US could be writing its own losing ticket’ in the AI race. Tom’s Hardware says projects worth $68 billion were reportedly blocked in the second quarter.
CFTC seeks comment on leveraged spot crypto trading for retail customers
The Commodity Futures Trading Commission published an advance notice of proposed rulemaking on retail crypto asset transactions and opened it for public comment. The Block reports that the framework would let crypto exchanges offer leveraged and margined spot trading to retail customers without action by Congress, and floats a new ‘crypto asset market’ exchange category. The agency said it cannot require crypto to trade on CFTC platforms without Congress. Chairman Michael Selig called the notice ‘a critical step’ toward rules designed to prevent schemes such as FTX.
WSJ: only 11% of surveyed businesses could forecast AI spending
The Wall Street Journal reports on a survey in which only 11 percent of 396 businesses said they could forecast their AI spending. AI use is billed in tokens, and the Journal says a model’s token use on a given task can be unpredictable. It also cites a Microsoft finding that lower-priced models ended up costing more than higher-priced ones on 32 percent of more than 6,800 tasks.
Quinnipiac poll: 77% of Americans want AI development slowed or stopped
A Quinnipiac University poll found that 47 percent of Americans want to reduce the pace of AI development and 30 percent want a full stop until safety has been verified, The Decoder reports. Five percent want faster development. Eighty-six percent support independent safety standards for AI companies, and 74 percent have little or no trust in the companies’ leaders. The poll surveyed 1,202 US adults from Sept. 24 to 27, with a margin of error of plus or minus 3.5 percentage points.
Senate blocks data center power-cost bill in 57–43 vote
The Ratepayer Protection Act fell three votes short of the 60 needed to advance in the US Senate on Sept. 30, Tom’s Hardware reports, citing Reuters. The bill, introduced in the Senate by Sen. Jon Husted, would have directed state regulators to consider whether large electricity users such as data centers should cover the incremental grid costs of serving them. The House had passed it two weeks earlier. Opponents called the bill ‘toothless’.
Reuters: Meta, TikTok and X challenge Ofcom’s information demands
Meta, TikTok and X are challenging UK regulator Ofcom over the amount of information it is demanding under the Online Safety Act, Reuters reports. The companies say the demands create unprecedented regulatory burdens.
Developer cloud Namespace raises $42 million Series B
Namespace raised a $42 million Series B led by Scale Venture Partners, Axios reports, and says it has more than 1,000 customers. Tech.eu reports the round comes seven months after its Series A, with participation from 20VC, NEA and Datadog chief executive Olivier Pomel. Founded by former Google principal engineer Hugo Santos, the company runs build and test infrastructure for software teams and AI coding agents.
Cerebras shares rise 9% after Altman calls it a ‘close partner’
Cerebras stock climbed 9 percent on Monday, CNBC reports. It had fallen 20 percent last week after it emerged that OpenAI would run the ‘Ultrafast’ mode for GPT-6.1 Sol on Nvidia GPUs rather than Cerebras chips. Sam Altman wrote on X on Friday that ‘Cerebras is a close partner’. CNBC puts Cerebras’s market value at about $43 billion, down from $95 billion after its May listing.
Ghost raises $11 million seed for a $3,499 personal AI computer
Ghost emerged from stealth on Monday with an $11 million seed round led by Andreessen Horowitz, TechCrunch reports. Founder Zain Javaid is 19. Its first product, Core, is a computer built to run AI agents continuously on a person’s own data; it costs $3,499 and includes an Nvidia RTX Pro 4000 SFF Blackwell GPU. Other investors include Abstract, Audacious Ventures, SV Angel and Nova.
arXiv limits researchers to two submissions a month
The preprint server arXiv will limit each researcher to two submissions per calendar month and three active submissions at a time, 404 Media reports. arXiv said it received 40,363 submissions in September 2026, against 20,569 in September 2024. Thomas Dietterich, chair of its editorial advisory council, said a relatively small proportion of authors submit large numbers of low-quality papers and take up a disproportionate share of volunteer moderators’ time.
Aleph Alpha releases open-weight German-English model Kolibri
Aleph Alpha released Kolibri, a German-English mixture-of-experts model with 78 billion parameters, of which about three billion are active per token, The Decoder reports. German makes up 21.3 percent of the training data. The model was trained on 768 B200 GPUs in Germany and Finland, and the weights are available under the Apache 2.0 license. Aleph Alpha says the model targets public administration, aviation and industry.
Florida woman charged after Anthropic reports Claude chat to police
A woman in Lee County, Florida, was arrested on a felony charge after Anthropic’s human review team reported a Claude conversation to law enforcement, Tom’s Hardware reports, citing WINK News. According to the arrest report, a Sept. 26 message said she would ‘shoot up’ the Lee County Sheriff’s Office; she described the chat as a diary. Court records list a Sept. 30 charge under Florida’s statute on written or electronic threats of a mass shooting. Tom’s Hardware says it is at least the third such conversation to reach police since August. Anthropic has not commented on the case.
Cohere launches North 2 enterprise agent platform
Cohere released North 2, a new version of its enterprise agent platform. VentureBeat reports that it adds cross-session memory and a redesigned harness. According to Cohere, the platform works with Cohere’s models or a customer’s own, and can be deployed in the cloud, on premises or in air-gapped environments. Administrators get cost controls, rate limits and user quotas.
Anthropic employees donated $540 million in 2025, IPO documents show
The Decoder, citing The Information’s review of documents Anthropic shared with prospective investors, reports that Anthropic employees donated $540 million in 2025. That is nearly five times the next-largest Fortune 500 donors, Truist Financial at $115 million and BlackRock at $109 million. The company tops up employees’ stock donations with additional shares. According to the documents, Dario Amodei and the six co-founders have pledged to give away at least 80 percent of their wealth. The Register’s Kettle podcast discussed the leaked prospectus, which the Financial Times reported lists existential risk from the technology.
Amazon hires Microsoft’s former Copilot CTO Pedram Rezaei
AWS hired Pedram Rezaei as a distinguished engineer in its Agentic AI and Emerging Technologies group, Business Insider reports, citing an internal email from vice president Swami Sivasubramanian. Rezaei spent two decades at Microsoft and was Copilot’s chief technology officer until June. He will initially work on the Quick assistant and the Kiro coding tool. The email says Quick’s monthly AI users grew fourfold from the first quarter to the second.
Bloomberg: tech billionaires account for all $845 billion in 2026 wealth gains
Bloomberg’s Billionaires Index shows tech billionaires account for all of the $845 billion in wealth gains so far in 2026. Billionaires whose fortunes came from outside tech lost a combined $62 billion. Bloomberg attributes the gains to the rise in AI valuations.
The Information: Jensen Huang’s son-in-law is a key Nvidia executive
The Information, citing sources, reports that Nico Caprez, Jensen Huang’s son-in-law, rose rapidly to become an Nvidia vice president. He serves as the company’s main contact for neoclouds and as Huang’s ‘consigliere’, the report says. Madison Huang, the chief executive’s daughter and an Nvidia executive, held her wedding in Hawaii a little over a week ago.
Nvidia raises Shield TV Pro price by $100 as memory costs climb
Nvidia raised the price of the Shield TV Pro from $199.99 to $299.99 on Oct. 2 because of higher component costs, The Verge reports, citing Ars Technica. The device launched nearly seven years ago. Tom’s Hardware reports that 32GB became the most common RAM configuration in Steam’s September survey at 42.22 percent, ahead of 16GB at 37.82 percent. The site also describes a repair shop fixing a pair of server DDR5 modules said to be worth $16,000, work that higher replacement prices now justify.
Researchers track AI agent fleet on Tencent infrastructure querying Amap
Independent researchers posted preliminary findings on Sunday about a group of AI agents that appear to run on Tencent’s infrastructure and target Alibaba’s map service Amap, TechCrunch reports. The agents were found through traffic to the domain-scanning service urlquery. Their queries sought directions to entrances of public places including a park, a zoo and a hospital. The researchers call it an ‘agent fleet’ rather than a swarm because they saw no sign of communication between the agents.
GitHub launches ReviewBench, an open benchmark for AI code review
GitHub released ReviewBench, a benchmark for code review agents. According to the GitHub Blog, it is built on representative GitHub pull requests and uses multi-source ground truth, calibrated evaluation and production-aligned metrics.
Security roundup: Korean bank breaches, Dell root flaw, 1,313 Debian CVEs
South Korea’s Financial Services Commission held an emergency meeting after cyberattacks on financial institutions and confirmed a data breach at Shinhan Bank, BleepingComputer reports; Yonhap reported that a server used in the attacks carried a page title associated with the open-source penetration-testing system ARTEX AI. Dell told customers to patch CVE-2026-86360 in its System Update tool, which allows code execution with root privileges. Italy’s data protection authority fined IQVIA €7 million over inadequate anonymization of about one million patients’ records. The US Justice Department announced the arrest of the alleged developer of the Ploutus ATM malware. The Register notes that Debian’s latest kernel security update lists 1,313 CVE identifiers.
On video: a16z on who pays for consumer AI and specialized models
Andreessen Horowitz posted two discussions. In one, partners Olivia Moore and Josh Elman go through the seventh edition of the firm’s Top 100 Consumer AI Apps; the video description says only a small share of consumers pay for AI and that spending is concentrated among power users. In the other, OpenRouter’s Alex Atallah and Replit chief executive Amjad Masad argue that specialized models routed by task can be cheaper and easier to control than one general model; the episode includes a segment on Stripe’s acquisition of OpenRouter. Separately, according to a generated summary of an All-In clip, Chamath Palihapitiya said companies deploying advanced AI need end-to-end traceability, policies mapped to risks and auditable evidence.