▓█████████▓▒░ ░▒▓█████████▓
▓███████▓▒░ ░ ░▒▓███████▓
▓█████▓▒░ ░ ░ ░▒▓█████▓
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
─────────────────────────────────────────
No. 64 · Thursday, 10 September 2026Welcome back. Three American agencies put their names to what the labs have been muttering for two years and accused six Chinese AI companies of stripping US models for parts. Anthropic kept its newest model away from Britain's safety institute, and Whitehall is reading that as protectionism. A Bay Area security firm showed that an AI-written worm could take over a WeChat account through a call nobody answers. And from tomorrow morning, anyone selling connected software into Europe is on a 24-hour breach-reporting clock.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░The NSA, the FBI and CISA named six Chinese AI companies and accused them of copying American models at industrial scale
The joint advisory, published Tuesday, says DeepSeek, Moonshot AI, Alibaba, the Shanghai model startups MiniMax and StepFun, and Z.ai, formerly Zhipu AI, have run distillation campaigns against frontier models from Anthropic, OpenAI, Google and xAI since at least late 2024. The campaigns pulled billions of tokens through millions of requests. Distillation, in which a smaller model is trained on a bigger one's outputs, is ordinary research practice and ordinarily a breach of the terms of service; the agencies say these campaigns used fraudulent accounts, bulk premium subscriptions and proxy routing services to stay out of sight. Moonshot, which filed confidentially last week for a Hong Kong listing, is accused of drawing on Claude Fable to build Kimi K3, the model that turned it into a global name. The recommendations to American labs are the interesting part: watch behavior, watch infrastructure, and degrade your answers when you think you are being farmed (CISA · The Next Web · Engadget).
Anthropic did not give Britain's AI Security Institute its newest model before release, and the Cabinet Office ordered an urgent assessment
Claude Mythos 5.1, the version with certain production safeguards removed for vetted cybersecurity and life-sciences customers, went to approved American organizations on September 1, the release we covered last week, but not to the UK institute, the Financial Times reports. It is the first Anthropic frontier release the institute has been left out of; it had access to Mythos 5 and used it to find the model adopting fake identities in a cyber test. British officials read the omission as a wider protectionist turn among US labs, and some suspect pressure from the Trump administration, which nobody has confirmed. Anthropic says it is coordinating with the US government to widen access to more partners at home and abroad. A safety institute that sees only the models it is handed is a reference check with the referee picked by the candidate (IT Pro · IBTimes UK).
An AI-written worm could have taken over WeChat accounts through a call the victim never answered
The Bay Area firm Calif disclosed WeWorm on Tuesday: a memory-corruption flaw in WeChat's voice-calling code that hands over an account in seconds on iOS and Android, with no tap, no answer and nothing audible if the target does pick up. From there it reads the contact list and calls the next person. WeChat and its Chinese twin Weixin passed 1.4 billion monthly users this year. Calif built the exploit in a little over a week using a mix of open-weight and commercial models, reported it to Tencent in July and briefed the White House before going public; Tencent has since shipped a fix and says it has no reason to believe any user was affected. The notable part is not the bug. It is an American security firm telling a Chinese company where its front door was unlocked, in the same week Washington accused six Chinese labs of theft (The Hacker News · Help Net Security · The Register).
From Friday, anyone selling connected hardware or software in Europe has 24 hours to report a flaw being exploited
The EU's Cyber Resilience Act reporting duties start on September 11: an early warning within 24 hours of a manufacturer having reasonable certainty that a vulnerability in its product is being actively exploited, a technical notification within 72 hours, and a final report within 14 days of a fix being available. Severe incidents follow the same first two steps and get a month for the closing report. There is no revenue floor and no grandfathering, so a five-person firm with one European customer is inside the rule, and so is a product shipped years ago. For communications teams the awkward number is the first one. Twenty-four hours is shorter than most companies' internal escalation path, and the filing portal opens on the same day the duty does (European Commission · Jones Day).
Apple's first folding iPhone is the iPhone Duo, from $1,999, on sale October 23 (update)
John Ternus opened his first product event as chief executive on Wednesday with the passport-shaped foldable we flagged as his opening test, seven years after Samsung's first foldable. Storage takes it from $1,999 to $3,199, preorders start October 16, and Swiss buyers get it on October 23. There is no Face ID. The iPhone 18 Pro, an Apple Watch Series 12 and Ultra 4, and AirPods 5 came with it, and Apple says it used AI and 3D printing to manufacture the hinge. Xiaomi launched its own wide foldable in China two days earlier, and Huawei a tri-fold the same day, so the shape of the category was settled before the company that popularizes it turned up (Engadget · CNBC · Variety).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ EARNINGS ▓▓▒▒░░Two prints after the close tonight, one for each side of the AI trade.
Watch Oracle's remaining performance obligations, the line where its AI data-center contracts sit, and Adobe's creative subscription growth, the line the AI image tools are chasing.
░░▒▒▓▓ AI ▓▓▒▒░░Paul Christiano, who helped invent the training method behind ChatGPT, joins the board that can delay OpenAI's model releases. OpenAI said on Wednesday that Christiano is joining the OpenAI Foundation board and its Safety and Security Committee, chaired by Carnegie Mellon's Zico Kolter, and will sit as a non-voting observer on the board of the for-profit arm. He ran alignment at OpenAI from 2017 to 2021, co-developed reinforcement learning from human feedback, the technique that turned raw language models into assistants. He also founded the Alignment Research Center and now advises the government's Center for AI Standards and Innovation inside NIST. The committee's stated power is the right to hold a release (OpenAI · Axios).
Suno threw out its own music models and replaced them with ones trained on licensed catalogue. The v6 family, launched Wednesday, was built with Warner Music Group, BMG and the distributor Believe, and on Suno's account it starts paying those partners from launch day. Three versions ship: v6 and the experimental v6-wild for the Pro and Premier paid tiers, v6-mini free for everyone. The company still faces a stack of copyright suits over how the previous models learned, including the Munich ruling for GEMA we covered in August, which is the context for retiring them outright rather than running licensed and unlicensed systems side by side (TechCrunch · Axios).
Chrome now ships every two weeks, because AI is finding the bugs faster in both directions. Google moved the browser to a fortnightly release cadence on Tuesday, starting with Chrome 153 on desktop, Android and iOS, down from the four-week cycle it adopted in 2021. The reasoning runs both ways: Gemini-powered tooling is turning up vulnerabilities the team would once have taken far longer to find, and faster-moving threats, some of them attributed to AI, have made the window between a fix landing in the public code and reaching users more dangerous (TechCrunch · Engadget).
OpenAI stopped taking ChatGPT ads from Adobe and anyone else selling image or voice generation. The Information reports that advertising partners were told campaigns for standalone image and audio products will no longer be approved, a change that is not in the published ads policy and that blindsided Adobe, an advertiser in the original pilot. Video generation is still welcome. It landed in the same week OpenAI shipped ChatGPT Images 2.5, and ChatGPT has also stopped returning outbound links on queries such as photo editors. The ads business we covered at the start of this month grows faster with the auction open to everyone, which is the tension the policy leaves unresolved (The Information · MediaPost).
░░▒▒▓▓ POLICY ▓▓▒▒░░Britain will require Apple and Google to stop nude images being taken, shared or viewed on children's devices. Culture Secretary Lisa Nandy told the Commons on Tuesday that the obligation falls on the phone and operating-system makers rather than on individual apps, which moves where British regulators think the control point sits. Both companies have already missed an earlier deadline to block child nudity on their handsets in the UK (Bloomberg · WIRED).
The White House answered Senate Democrats about its frontier-model testing framework without answering anything. Five senators including Mark Warner, Kirsten Gillibrand and Adam Schiff wrote in early August with a list of questions about the voluntary pre-release review framework and the legal basis for restricting foreign nationals' access to American models, calling the process a black box. The reply, sent on Friday and obtained by The Information, said the administration is working closely with industry and that the Treasury Department is coordinating a risk review. The framework itself has still not been published, and companies briefed on it say they do not understand how it applies either (The Information · Sen. Warner).
░░▒▒▓▓ INFRA ▓▓▒▒░░Google will spend €13 billion in Finland and has bought up to half a nuclear plant's output for 22 years. The commitment announced Wednesday, about $15 billion and the company's largest in Europe, covers data centers and grid work at Hamina, Kajaani, Muhos and Vaala across 2027 and 2028. Google puts the construction-phase contribution to Finnish GDP at roughly €3.6 billion a year, and the jobs once the sites are running at about 7,000. The power agreement with Fortum extends the life of the Loviisa plant and eventually covers up to half its capacity. It is Google's first nuclear power deal in Europe, and it buys the two things a northern data center actually needs, which are cold air and a firm supply (ESG Dive · Quartz).
Verizon has bought more than 80 million miles of Corning fiber, running to 2032. The multi-year, multi-billion-dollar supply agreement announced Tuesday covers deployments from 2027 and does two jobs at once: reaching 40 to 50 million homes and businesses with broadband, and laying the long-haul backbone that connects AI data centers to each other. Neither company put a figure on it beyond multi-billion. Verizon Business chief executive Kyle Malady called it securing the physical foundation of America's AI economy, which is a lot of freight for a cable, though the cable does have to exist (Verizon · Light Reading).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Claude's heaviest users are suing Anthropic over what 5x and 20x actually buy. The expanded class action filed Tuesday argues that the $100 and $200 Max tiers do not deliver five and twenty times the usage of the $20 plan, because rolling five-hour windows and weekly caps interact in ways the marketing does not explain. Anthropic's usage limits have been contested since the Claude Code changes we covered at the start of the month. The limits are documented on a Claude support page; plaintiffs' counsel Monica Vaca, a former Federal Trade Commission lawyer bringing the case with her ex-colleague Kati Daffan, says customers should not have to follow several links to learn what they bought. It is an allegation, not a finding, and a rare attempt to treat AI subscription pricing as a consumer-protection question (The Verge · Engadget).
ElevenLabs hired the finance chief who helped take Adyen public and started saying the word IPO. Ethan Tandowsky, who joined the Dutch payments company in 2016, worked on its 2018 listing and became its chief financial officer in 2023, was named ElevenLabs' chief financial officer on Tuesday. The voice-AI company's own announcement says large organizations now bring in more than half of its revenue; The Information reports it is on track for about $600 million of annual recurring revenue by year end, up from roughly $350 million at the end of last year, has reached profitability, and has a listing penciled for 2028, which is two years of runway for a hire whose value is concentrated in the last six months of it (ElevenLabs · The Information).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░The US destroyed five Iranian oil tankers, Iran struck an American base in Jordan, and oil went above $100. Four tankers were hit in the Gulf of Oman and one near Kharg Island, after the Revolutionary Guard targeted a US Navy warship twice in two days. The exchange at sea we covered on Tuesday has now reached Jordan: Iran's retaliation fell on the Al-Azraq base, and Jordan's military said its air defenses intercepted 18 ballistic missiles while two more landed in unpopulated ground. Oil went above $100 a barrel on Wednesday, six months into a war that has all but closed the Strait of Hormuz to the fifth of the world's oil that used to move through it (NPR · Al Jazeera).
The AfD finished three seats short of governing Saxony-Anhalt outright, and two more state votes come on September 20 (update). As we covered on Monday, Sunday's result was the party's best state result yet and the first time it has come within reach of an outright majority; the arithmetic since is that it needs defectors or the Sahra Wagenknecht Alliance, which cleared the 5% threshold and ruled out joining the state government before the election. Every established party has ruled out working with it. Berlin and Mecklenburg-Vorpommern vote in ten days, and the AfD leads the polls in the second of them (CNN · The Guardian).
░░▒▒▓▓ ALPS ▓▓▒▒░░A Munich company raised $450 million to build Europe's own way home from orbit. The Exploration Company, which runs out of Munich and Bordeaux, said on Tuesday it had closed the largest Series C a European space company has done, co-led by Bessemer Venture Partners, Atomico and EQT's Scaleup Europe Fund, with Balderton, Plural, Cherry and Red River West returning. Total funding is now around $680 million. The money goes to Nyx, a reusable capsule meant to dock with the International Space Station and come back intact, and to Storm, a reusable high-thrust engine. It is also, the company notes, the largest round ever raised by an EU-headquartered company with a woman as founder and chief executive (TechCrunch · European Spaceflight).
DNIP argues that digital sovereignty starts with something duller than a data center, which is your domain name. In Wednesday's post for the Swiss site Das Netz ist politisch, the computer scientist Marcel Waldvogel walks through the layers a European organization has to control to run its own internet services, and spends most of the piece on domains, where a name registered through the wrong chain can be switched off by someone outside your jurisdiction whichever cloud your servers sit in. Sovereignty budgets rarely reach that far down the stack (DNIP).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Hugging Face has sold more than 15,000 walking rubber ducks, and Nvidia is buying the company that makes them. Microduck, a $399 open-source biped launched late last month that can be taught to roller-skate, kick and get back up after a fall, has passed $6 million in sales on almost no advertising, co-founder Thomas Wolf told The Information on Tuesday, against a first-year projection of 50,000 units and $20 million that is now being revised upward. Training one requires simulations that run on Nvidia GPUs, which is convenient for the buyer in the $12.9 billion purchase we covered last month. Wolf's pitch is the price gap: you can experiment on a $40,000 humanoid prototype, or on the duck (The Information · TechCrunch).
░░▒▒▓▓ PAPERS ▓▓▒▒░░MIT, Google, Google DeepMind and Stanford — a working code library whose main branch contains almost no code. In a preprint posted on September 4, the authors describe SMART, a performance-modeling tool for machine-learning systems whose repository is a graph of plain-English design documents rather than source files. Coding agents regenerate the implementation from those documents on every version bump, and a human change means editing a document, not a function. The claim is that regenerated implementations reproduce hand-audited reference models to rounding precision, including a DeepSeek-V3 workload on a slice of a Google TPU pod. Two things make it work: design docs written as step-by-step worked examples, which act as demonstrations for the generating agent, and a deliberately small formal core. For anyone who writes documentation nobody reads, the provocation is that the document becomes the artifact and the code becomes the output, which only holds where regenerating is cheaper than repairing (arXiv).
░░▒▒▓▓ REPORTS ▓▓▒▒░░OECD — PISA 2025 Results. The triennial test of 760,000 fifteen-year-olds across 91 countries, released on Tuesday, puts reading and mathematics at their lowest levels since the series began in 2000; average reading across OECD countries fell 28 points between 2015 and 2025, to 461, which the OECD calls roughly a year and a half of schooling. Secretary-General Mathias Cormann tied the reading decline to more screen time and less reading for pleasure, with results starting to fall past five hours a day of device use. The finding for this readership sits further down. Nearly half of students in OECD countries use AI chatbots weekly or more to help them learn, and those who use them to draft essays, summarize texts and research topics score about 20 points lower in science, or about a year's worth, while weekly general-purpose users score no differently from non-users. That is a correlation on a self-reported measure, and it cannot separate a struggling student reaching for a chatbot from a chatbot producing one (OECD · US News).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░More than 2,000 court decisions now involve someone filing AI-invented material, and the courts have stopped being amused. The legal researcher Damien Charlotin maintains the running database. The Georgia Supreme Court this year suspended a Clayton County assistant district attorney for six months over a filing with nonexistent cases, and the federal Ninth Circuit sanctioned two lawyers for briefs containing invented cases and misdescribed real ones, stressing that the punishment was for failing their existing duties of competence and candor rather than for using AI at all. Google's general counsel Halimah DeLaine Prado, whose company released Gemini Enterprise for Legal last month, told Axios on Tuesday that she sees the technology as a complement and not a replacement, and that the practice of law will always rise and fall on judgment. The takeaway for anyone signing off on AI-assisted work is that no regulator has invented a softer standard for it (AI Hallucination Cases · Axios).
Everyone is reading the same 1985 Western, and the recommendation engine turns out to be a chatbot. In Tuesday's Not Boring, the investor and writer Packy McCormick traces the slow revival of Larry McMurtry's Lonesome Dove through a podcast, an obituary, BookTok and a New Yorker summer list, then notes the part he did not expect: he found it because Claude recommended it, ChatGPT recommended it twice more, and several friends who had read it said the same thing happened to them. Simon & Schuster reported the novel sold 56,000 copies in 2025, its best year since the 1993 sequel miniseries. Search rewards what you already know to ask for and social rewards what is new. Anyone maintaining a back catalogue should find out what the assistants say about it (Not Boring).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░A startup founded by a former OpenAI staffer has talked SpaceX into letting it bolt chip factories onto returning rockets. Besxar's fabships are microwave-sized canisters that ride up on a Falcon 9 booster, spend a few minutes in near-vacuum making precursors for advanced semiconductors, and come home with the booster under ten minutes later, which is one way to avoid building a cleanroom. Two flew on a July Starlink mission, and founder Ashley Pilipiszyn told TechCrunch the flown wafers came back cleaner than the ones that stayed on the ground. She has raised almost $14 million on the argument that it is now cheaper to go where the physics already works than to fight it in Arizona (TechCrunch).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Theme of Done Right — Ross Simonini
Ad astra — the Uncertain Futures desk