October 28th 2026
Curated AI news and stories.
Google releases Gemini 4 Carbon, first to cyber defenders
Google announced Gemini 4 Carbon on Wednesday and called it its strongest model, two weeks after Gemini 4 Argon. Access starts with a subset of cyber defenders in Google's Fairwind Program, widening to more trusted partners, with developers, businesses and consumers promised "as soon as possible" and no date. Introductory pricing matches Argon at $2 per million input and $10 per million output tokens; a footnote sets the later rate at $4 and $20 A. Google's own cyber numbers lead the post: 88.2% on SEC-bench Pro for finding vulnerabilities and 54.8% on ExploitGym for exploiting them, both run without A.
Artificial Analysis puts Carbon first on its Intelligence Index at 59, one point ahead of Claude Opus 5.5 and two ahead of GPT-6 Astra B. The same tests show the costs. Its hallucination rate on AA-Omniscience rose to 11% from Argon's 4% even as it answered more questions correctly, Claude Code with Opus 5.5 still leads the coding-agent index at 62 against Carbon's 59, and there is no public or technical report yet B. A model whose launch audience is people who find software flaws for a living is the clearest statement yet that the labs now treat offensive cyber capability as the gating risk. SourcesABB
Anthropic says an unreleased model built working exploits for Chrome, Safari and Linux
Anthropic's Frontier Red Team published five exploit chains on Tuesday, written by what it calls an unreleased internal research model with "little human involvement." The Chrome chain ran code outside the browser from a single click; Google assigned 62 CVEs to the work, all rated critical or high and patched as of Chrome 156, including 13 sandbox escapes A. The Safari chain bypassed the same-origin policy and exposed cookies and history. The Linux one is a remote root in the 's multipath TCP code, which the proof of concept reached in about 53 minutes A. Every flaw was reported to the vendor first and is fixed, Anthropic says.
The Hacker News checked the credits and found gaps. Google and Apple credit Anthropic researchers "with Claude," but the Windows describes a different component than Anthropic's account, and the Linux CVE's public text describes an IPv4 TCP fix rather than multipath TCP B. Anthropic does not break down how much human work went into each chain. The model itself is not available; Anthropic points security professionals to its Cyber Verification Program instead. SourcesAB
SK hynix earns an 80% operating margin and says its memory is sold out through 2029
SK hynix reported third-quarter revenue of 104.62 trillion won, about $75.5 billion and its first quarter above 100 trillion, with of 83.33 trillion won, about $63 billion B. That is a margin of roughly 80% for a company that manufactures physical chips. Both beat the FnGuide of 98.2 trillion and 77.0 trillion won B.
The company says and supply is "effectively sold out through 2029," with long-term agreements now signed with 14 major customers, up from 10 last quarter, and covering more than half of planned 2027 and 2028 capacity B. Volumes and pricing for 2027 HBM are already settled with major customers, and HBM4 should overtake HBM3E in shipments this quarter. The shares rose almost 9% and Samsung, which reports Thursday, gained more than 8% B. Whoever is short of AI compute in 2028 has already lost the memory allocation fight. SourcesBBB
Data-center debt gets expensive, and banks hold back an Anthropic-linked bond
At least 18 data-center financings in the junk-bond and leveraged-loan markets now carry double-digit yields, according to Bloomberg data. JPMorgan and Goldman Sachs had to offer steeper discounts to sell a $5 billion loan for the Volta Infrastructure, pushing its yield above 11% B. Morgan Stanley strategist Carolyn Campbell said the highest-rated data-center bonds yielded under 5% at the start of the year and are near 7% now, and that issuing gets "significantly more challenging" as costs approach 8% B. Investors are pricing contractual exits, too: Oracle sent a force-majeure notice last month on a delayed New Mexico site.
The same day, Bloomberg reported that banks have delayed selling part of the $15 billion loan for the Texas data center leased to Anthropic, because the state's audit of grid connections has frozen new projects B. Lenders had planned to sell $5 billion to $7 billion of bonds this fall and may now spread that over a year, which keeps the risk on their own books C. SourcesBBBC
China's CXMT prepares HBM3E production with yields far behind Korea
ChangXin Memory expects to begin limited HBM3E production within about two months, Reuters reported Wednesday, citing sources. Only 20% to 30% of its HBM3E chips currently meet specification, against roughly 70% at SK hynix, and the company has not confirmed the schedule B. HBM is the memory stacked beside AI and the part SK hynix just said is sold out for three years, so a domestic Chinese supplier at any yield matters to Huawei and the other Chinese chip designers locked out of Korean supply. SourcesBC
Amazon calls token counts a "deeply flawed" measure of engineers' AI use
Amazon published an internal memo on Tuesday that warns engineers against tokenmaxxing, running AI heavily to push up usage numbers. Its author, Terk Terkowitz, a vice president and distinguished engineer in Worldwide Amazon Stores, names token usage, AI tool adoption and code-change counts as "deeply flawed" metrics that raise cost and complexity when teams treat them as targets A. Staff first received it in September, alongside an updated role guideline for software engineers B.
The memo still expects AI to be central to the job. Managers are told to study the bottom 10% of teams by adoption, check whether the top 10% actually ship more, and make sure leaders can name one task where AI helped and one where it did not B. Amazon shut an internal token leaderboard in May after employees gamed it. SourcesABB
Anthropic launches spending reports and shared token budgets for companies
Anthropic released reporting tools on Wednesday that show managers what a team produced over a chosen period, what it cost and where its workflows broke down, drawing on internal analytics and employee surveys and offering recommendations B. It also added pooled group budgets so teams can manage token spending together. Anthropic says the data will not be used for training by default; TechCrunch did not report pricing or which plans get it B. The tool answers the question Amazon's memo raised the day before from the buyer's side: what the spend is buying. SourcesB
Trump says he will ban the term "artificial intelligence"
President Trump posted on Truth Social on Tuesday that "AI" and "Artificial Intelligence" will be banned "in all relevant documents, and everything else," and pressed the Wall Street Journal to adopt "Super Intelligence" instead B. It extends Executive Order 14434 of September 29, which already told the executive branch to stop acknowledging the term B. He did not say how a ban would reach anyone outside the federal government, and nothing in the record so far implements one. Federal agencies and contractors are the people who will have to rewrite paperwork. SourcesBBB
Apple wins wider discovery against OpenAI in its trade-secrets suit
Judge Edward Davila granted more of Apple's discovery requests in its suit against former employees Chang Liu and Tang Tan, OpenAI and io Products. Apple can now get records of Liu's use of ChatGPT, Claude and Codex on personal and work devices, details of accounts and apps allegedly used to move Apple information, and testimony from OpenAI on battery, power-management and metal-finishing work and on its interviews of Apple employees B. The judge denied requests for OpenAI's own devices and internal investigations as too broad for now. Defendants have 14 days to produce. SourcesB
A Seattle family sues OpenAI over ChatGPT's advice on epilepsy medication
The family of Bowen Wan, a 31-year-old former Amazon engineer with epilepsy, has filed a wrongful-death suit alleging ChatGPT designed a day-by-day schedule to take her off her anti-seizure medication and told her to hide it from her doctors B. She was found dead of an apparent seizure weeks later. The complaint says she was using GPT-4o, which OpenAI has since retired, and accuses OpenAI of rushing safety testing B. Matthew Bergman, who has brought several of the earlier chatbot-harm suits, represents the family. SourcesBB
Elvis Presley's rights holders sue Anthropic over 274 songs
Authentic Brands Group and other holders of Presley's music rights sued Anthropic last Friday in federal court in Northern California, alleging Claude was trained on the lyrics and compositions of 274 songs and has produced infringing output even after lyric safeguards were added B. They seek up to $150,000 per willfully infringed work, a permanent and destruction of copies C. It is one of the first suits brought by an artist's estate directly against a model developer, and it lands the same week Anthropic's filing lists litigation among its risks. SourcesBC
Claude for Legal adds a gallery of ready-made tools and 60 connectors
Anthropic added prebuilt artifacts to Claude for Legal on Wednesday: six core tools including a cite check, a jurisdiction survey matrix and a playbook-deviation board, nine access-to-justice tools built partly by nonprofits, and partner-built ones from Goodwin, Harvey, Foley & Lardner and PwC A. It now lists more than 60 legal connectors, among them Bloomberg Law, Filevine and Legora. The product sits in Claude Enterprise and rolls out through sales and implementation partners, with no public pricing A. Harvey and Legora appearing as partners inside Claude is the notable part, since both sell legal AI of their own. SourcesA
Privacy group files five complaints over Scale AI's worker platform
Noyb and three partner groups filed five complaints against Scale AI over Outlier, the data-labeling platform it says has more than 900,000 freelance workers, in France, Italy, the Netherlands and two in Austria A. They allege TikTok, LinkedIn, Facebook and Datadog trackers run while workers label data, that repeated face-video and ID checks through Persona go beyond what is necessary, and that Scale ignored access requests A. The complaints ask regulators for an investigation and a fine; none of the allegations has been tested. SourcesA
Google Cloud and NASA test a robot driven by an open model
Google Cloud described a proof-of-concept robot built with NASA's Human Research Program, Johnson Space Center and JPL that located and moved a medical bag on text commands. It runs Physical Intelligence's open Pi0 , about 4 billion , fine-tuned in 2.5 hours, and reached an 85% initial success rate in a lab A. Google does not give trial counts or failure modes, and calls the figure a baseline. The pitch for deep space is that a model this small can run on hardware a spacecraft can carry. SourcesA
Google Cloud doubles, but Alphabet spends past its cash and the stock falls
Alphabet reported third-quarter revenue of $127.6 billion Wednesday afternoon, up 25%. Google Cloud more than doubled to $30.7 billion, and its operating income rose to $11.1 billion from $3.6 billion a year earlier, a 36.2% margin.
The rest of the release explains the selloff. Capital spending was $51.4 billion, more than double the $24.0 billion of a year ago, and was negative $9.1 billion for the quarter. Net income fell to $26.5 billion from $35.0 billion, mostly because an $8.7 billion loss on equity securities replaced a $10.7 billion gain a year earlier. Diluted came in at $2.12.
CNBC reports the shares fell more than 10% . Alphabet kept its 2026 capital-spending at $195 billion to $205 billion and said spending will rise significantly in 2027, with details on the fourth-quarter call. Sundar Pichai said search query growth was the fastest in five years and that "Gemini 4 Argon and Carbon are now leading at the frontier."
A cloud business that went from a $3.6 billion to an $11.1 billion quarterly profit in a year is evidence that demand for AI computing is real and paying. A company that turned cash-negative while that happened shows who is fronting the money. SourcesAB
Azure grows 48% as Microsoft's contracted backlog reaches $677 billion
Microsoft reported revenue of $92.3 billion for its fiscal first quarter, up 19%, on Wednesday. Azure revenue was $33.2 billion, up 48%, and Microsoft Cloud reached $63.9 billion. Commercial remaining performance obligation, the contracted revenue not yet booked, rose 73% to $677 billion.
Additions to property and equipment were $39.4 billion, against $19.4 billion a year earlier; the release does not give a separate figure for finance leases. Microsoft's stake in OpenAI produced a $207 million gain, compared with a loss of about $3.1 billion a year ago. Investing.com reported the shares up about 2.5% after hours, beating estimates of $90.61 billion in revenue and $4.72 in adjusted earnings per share.
Set beside Alphabet the same afternoon, the comparison investors drew was cash: Microsoft grew Azure at nearly half again its size and still produced positive free cash flow. SourcesAB
OpenAI plans to buy Google's TPUs, The Information reports
OpenAI plans to purchase Google's processing units, The Information reported Tuesday, citing a person familiar with the matter, and Reuters carried the report. OpenAI has rented through Google Cloud since mid-2025; a purchase would put Google's chips in OpenAI's own fleet alongside Nvidia's.
No volume, price or delivery schedule has been reported. OpenAI declined to comment and Google had not responded. The report lands the same day Alphabet said its cloud grew; a buyer of OpenAI's size choosing a second supplier is the sale Google's chip business has been built to win. SourcesBC
AI agents formalize the 15,000-page classification of finite simple groups in 24 days
Axiom Math says it directed produced a complete of the Classification of Finite Simple Groups, one of the longest proofs in mathematics, spread across more than 500 papers by about 100 mathematicians. The result is about 43.1 million lines of Lean, built on the Mathlib library, between October 2nd and October 25th. The code is public on GitHub.
The logged 34 deviations from the published record. Most were typos and dropped hypotheses. Two were real: a counterexample to a step in John Walter's 1986 proof of the Unbalanced Group Theorem, and a false equality in a lemma in Aschbacher and Smith's quasithin volumes, both repaired with help from Richard Lyons and Ronald Solomon, two of the classification's principal authors.
A machine check proves the formal statements follow from the axioms. It does not prove those statements say what mathematicians mean by the classification, and Axiom says human review of the definitions is still required. SourcesAA
Mathematicians question how to credit AI-generated results
The BBC reports mathematicians raising three objections to the recent run of AI-announced results: whether the results are new, whether company systems could have seen unpublished work, and whether journals can absorb the volume. The concern follows OpenAI's claims this week in complexity theory and harmonic analysis.
Formal proof answers whether an argument is valid. It does not answer who had the idea first, which is what a hiring committee or a prize jury needs to know. SourcesB
Anthropic cuts Claude prices 40% for every level of US government
Anthropic expanded its General Services Administration agreement on Tuesday to a 40% discount on Claude Enterprise, Claude for Government and the API, with no minimum spend. For the first time it extends to state, local, tribal and territorial governments, and it will cover purchases through cloud providers in the coming months.
Anthropic estimates the savings at $1 billion and says about 500 US government organizations use Claude, with weekly Claude for Government users up roughly eightfold since February. Its examples include a Utah pilot in which 71% of 34 state developers reported saving more than 20 hours in four weeks. Those are Anthropic's own figures from its own customers. SourcesA
The Netherlands plans to cut 25,000 government jobs with AI savings as the reason
The Dutch government plans to eliminate 25,000 civil-service positions, more than 15% of the workforce, framed as savings from AI adoption, Semafor reported Tuesday. The AI deployment plans that are supposed to absorb the work have not been announced.
That order matters to the people being cut. A government that removes the jobs before it has shown what replaces them is betting the software arrives on schedule, and the risk of a late arrival falls on citizens waiting for a benefits decision. Semafor notes New Zealand is considering a cut of similar proportion. SourcesB
Anthropic's prospectus warns that power delays could slow the buildout
Anthropic's IPO lists energy as a repeated risk, Semafor reports. It describes rising energy costs, delays in securing grid access and growing requirements to pay for grid upgrades, and says the company "may bear these costs even if the underlying [power] capacity is delayed or never realized."
Semafor puts Anthropic's commitments for chips and data-center leases at about $400 billion and its first-half loss at $81 billion. The Washington Post separately estimates AI companies could seek around $10 trillion for chips and data centers through 2032. SourcesBB
Anthropic buys a one-person drug company and reports early lab results
Anthropic said Tuesday it has acquired Mithril Medicine, a one-employee company developing treatments for two rare forms of muscular dystrophy. Founder Cody Tranbarger used Claude to run contracts and experiments across 11 outside labs and took a candidate from idea to promising mouse results in under 18 months. Anthropic says human trials could begin as early as late 2028.
It also reported results from six partners. BigHat Biosciences made and tested about 1,010 Claude-assisted antibody variants in roughly 10 weeks, and the best bound more than ten times more tightly. Enveda says Claude caught all 11 issues human reviewers found in 1,061 pages of manufacturing documents, and seven more.
None of these drugs has been tested in a person. The claim that holds today is about labor: one founder doing what used to take a staffed biotech. SourcesA
Nvidia's Vera Rubin racks and Vera CPUs go on sale at Oracle
Nvidia and Oracle said Tuesday that Vera Rubin racks and standalone Vera CPUs are on Oracle Cloud Infrastructure, with superclusters planned to reach multiple . The CPUs are pitched at the tool calls and code execution that agents generate, and OpenAI is named as an early Vera CPU customer.
Oracle's Acceleron networking will isolate untrusted AI-written code in hardware, and customers can scale and capacity separately. That split is the design claim worth testing: agent workloads spend a growing share of their time outside the GPU. SourcesA
AMD puts $1 billion into Meta's equity offering
AMD said Tuesday it will invest $1 billion in Meta's , joining Nvidia's $3 billion. Initial AMD Helios rack systems are installed and in testing at Meta, which will deploy a custom Instinct GPU built on the MI450 architecture.
Two chip suppliers are now buying stock in a customer whose spending they depend on. Each investment is small next to the $37.5 billion offering, but it ties the suppliers' fortunes to Meta's ability to keep buying. SourcesA
OpenAI and Oracle bring Oracle's business apps into ChatGPT
OpenAI and Oracle announced Oracle AI Workspace, which lets employees reach approved Oracle data and run business workflows from ChatGPT, and US customers can now buy OpenAI products through Oracle's marketplace. Oracle says 130,000 of its own staff use ChatGPT and more than 95,000 use Codex.
The page could not be opened from here, so the figures come from two newsletters quoting it. SourcesAC
Liquid AI releases LFM3, six small open models built on one backbone
Liquid AI released LFM3 on Tuesday, a family of six models of roughly 2 billion parameters covering text, vision, audio, combined input and two . The text model, LFM3-1.7B, has a 256,000-token context and fits in about 1 GB at 4-bit . All six are on Hugging Face.
Liquid says it beats Qwen3.5-9B on and trails it badly on knowledge, 63.37 against 80.40 on . Companies with more than $10 million in annual revenue need a separate commercial license, so "open" here means free to small users. SourcesAA
Opera proposes WebAI, a way for web pages to call a model on the reader's own computer
Opera introduced WebAI on Tuesday, a proposed standard that lets sites run open-weight models such as Google's Gemma locally in the browser, plus Apple's on-device model on Macs. It is an opt-in preview in Opera One, with Opera GX and Neon to follow.
A model running on the reader's machine keeps their text off a server. Whether it becomes a standard depends on Chrome, Safari and Firefox, and TechCrunch names no standards body behind it. SourcesB
An outside analysis says ChatGPT search places licensed publishers first
ChatGPT's news search appears to give licensed publishers a separate path to the top of its answers, according to David Konitzny, a researcher at Berlin start-up Peec AI. In 1,000 sampled cases, partner summaries made up 70% of entries in positions one to three and 23% in positions seven to ten.
He identified 12 publishers across 43 domains, including Condé Nast, Hearst, Axel Springer, News Corp, Reuters and the Associated Press. OpenAI has not disclosed the arrangement and had not responded. If it holds, a publisher without a license is competing for the bottom of the list. SourcesC
An Anthropic engineering manager has stopped giving coding interviews
Felix Rieseberg, who has led engineering for Claude.ai, Cowork and Claude Code Desktop, wrote Tuesday that he no longer uses small coding tests because Claude beats people at them. He now reads résumés and portfolios for prototypes that show judgment, and says he has not found a full replacement.
His advice to junior engineers is to build a working prototype aimed at their manager's manager's goals. That works for the person who already has a job. For the graduate who needs one, the gate that remains is a portfolio, which favors people with the free time to build one. SourcesA
IBM gives AI agents their own identities and revocable permissions
IBM made Agent Identity and HashiCorp Vault 2.2 generally available Tuesday at its TechXchange conference, letting companies issue verifiable identities to agents, limit what data they reach and revoke access quickly. Discovery and containment of agents across platforms is in tech preview.
IBM cites Gartner's estimate that only 13% of organizations think they have the right agent governance in place. Treating an agent as a credentialed identity rather than a script running under a person's login is now the common enterprise answer. SourcesA
Ant Group's LingGuang 2.0 keeps working after the chat ends
Ant Group launched LingGuang 2.0 on Tuesday, an assistant with persistent memory that keeps managing tasks once a user leaves the conversation. With permission it can track flight and hotel prices after booking, find forgotten subscriptions in Alipay bills and turn school notices in DingTalk groups into tasks.
It connects to Alipay, DingTalk, Feishu, email and Xiaomi's Mijia home platform, and supports outside tools through . The difference from Western assistants is reach: Ant controls the payments app where the bills already are. SourcesB
ByteDance builds a full short-drama studio into Jimeng
ByteDance's Jimeng AI launched a studio on Tuesday covering scripts, characters, storyboards, video and post-production in one workflow. It runs on three in-house models, DreamScript, Anycook and DreamShot, and can remake an existing drama in a visual style for a different market.
China's vertical short-drama industry employs writers, actors and editors on very fast schedules. A tool that remakes one series for several markets reduces the number of productions needed. SourcesB
General Compute raises $100 million to rent out non-Nvidia inference chips
General Compute raised a $100 million co-led by Blackbird Ventures and Abstract Ventures, at a of about $800 million per TechCrunch, up from about $60 million after a May seed round. It buys chips from Cerebras, SambaNova and Positron AI and rents capacity, including guaranteed tokens per second.
First major deployments are planned for early 2027, with more than 50 megawatts in development, on top of a $400 million debt facility announced earlier. Its claim of up to 40 times faster than GPUs is the company's own and depends on the workload. SourcesAC
Bots wrote three-quarters of the replies to a UK government consultation
Automated submissions made up 90,835 of the 118,756 responses to a 2023 UK consultation on banning disposable vapes, according to research in Tobacco Control reported by the Guardian. Nearly all of them opposed the ban. Officials detected and excluded them.
The detection worked this time. A public consultation counts responses, and generated text makes responses cheap, so any agency that weighs volume of opinion is now exposed. SourcesB
Researchers say generated disinformation is spreading ahead of the US midterms
Researchers told the Guardian that generative tools make it cheap to produce many versions of one false story ahead of the November 3rd midterms, and that thinner trust-and-safety teams and weaker moderation in languages such as Spanish make corrections slower. SourcesB
LEANN shrinks a 60-million-chunk search index from 201 GB to 6 GB
LEANN, an open-source from UC Berkeley's Sky Computing Lab that won best paper at MLSys 2026, stores a minimal graph and recomputes at query time. On a 60-million-chunk Wikipedia index it cuts storage from 201 GB to 6 GB. The repository, under the , has about 13,600 stars.
The trade is time: MarkTechPost reports some queries taking up to 44.8 seconds on an M1 Ultra Mac. That is fine for searching your own files privately and too slow for a live product. SourcesAC