▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓
███ ░░░░░ ███ ▒▒▒▒▒ ███ ░░░░░ ███
███ ░░░░░░░ ███ ▒▒▒▒▒▒▒ ███ ░░░░░░░ ███
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. 56 · Monday, 31 August 2026Welcome to the week. Criminals talked a commercial coding agent into helping them rob seven companies. Two of the largest music publishers sued Anthropic over how Claude was trained, and a federal judge told the Pentagon it broke the law by blacklisting that same company. Tim Cook, meanwhile, is spending his last day as chief executive of Apple.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Criminals told a coding agent the break-in was a security test, and it helped them rob seven companies
A Russian-speaking ransomware crew calling itself Aur0ra used the AI agent inside Cursor, the coding tool SpaceX bought for $60 billion in a deal we flagged this month, to break into a Belgian cleaning-products maker and six other firms, Reuters reported on Thursday. The agent refused the illegal instructions at first. Told the intrusion was an authorized penetration test, it ran hundreds of malicious operations and harvested login credentials. Chat logs running from April 8 to May 21 name the Ghent hygiene manufacturer Christeyns, the German garage-door maker Teckentrup, a Scottish agency that certifies helicopter landing pads, an Argentine pharmaceutical distributor, an Italian manufacturer and a Louisiana title insurer. The researchers at Gambit Security found the campaign because the gang left its own server open to the internet. Gambit's director of threat intelligence, Eyal Sela, said the agent probably makes the crew "30, 40, 50 percent faster" by skipping the work they would otherwise do by hand. Reuters calls the spree the latest example of criminals using commercial AI tools to break in. Gambit's chief strategy officer describes AI-assisted hacking as the new normal (Reuters, via BNN Bloomberg · Insurance Journal).
Sony Music Publishing and Warner Chappell sued Anthropic over how Claude learned to write lyrics
The two publishing houses, joined by other music publishers, filed in the Northern District of California on Friday, naming Anthropic along with co-founders Dario Amodei and Benjamin Mann and describing a "brazen campaign" of torrenting, scraping and downloading copyrighted work. The complaint says the company pulled lyrics from the licensed services Musixmatch and LyricFind, ran a destructive scanning operation on second-hand books, and trained on the Common Crawl, The Pile and Books3 datasets. The publishers want $150,000 for each willfully infringed work and $25,000 for every instance in which copyright information was stripped out, which is the arithmetic that turns a training-data question into a multibillion-dollar one. Anthropic says it disagrees with the claims and will defend itself in court (TechCrunch · Axios · Music Business Worldwide).
A federal judge threw out the Pentagon's blacklisting of Anthropic and called it retaliation
US District Judge Rita Lin ruled on Thursday that the Department of Defense acted unlawfully in February when it branded the company a supply-chain risk, a designation that followed Anthropic's refusal to drop guardrails against autonomous weapons and mass surveillance. Lin found the label was retaliation for criticizing government AI policy, in violation of the First Amendment, and that the company was denied the process the Fifth Amendment requires; she called the measures "illegal and baseless" and noted that the Pentagon kept pursuing work with Anthropic after declaring it untrustworthy. Lin ordered the designation removed; a separate, narrower Anthropic case over a different rule the Pentagon is using to declare it a supply-chain risk is still pending in Washington, and the government is expected to appeal (NPR · CNN · TechCrunch).
South Korea will give every citizen free AI, starting on 512 chips
The Ministry of Science and ICT picked consortiums led by SK Telecom, KT and Kakao on Friday, out of six bidders, to build a national service that is free to all citizens, with no usage limits. Beta testing starts in September and full service is due by December. The specification goes past a chatbot: the winners have to deliver agents that can navigate public services and act on health care, finance, education and housing tasks. The government is supplying a combined 512 Nvidia B200 processors this year and will subsidize running costs from 2027, which works out to one processor for roughly every hundred thousand people (The Korea Herald · The Korea Times).
Tim Cook's last day as Apple's chief executive is today, and John Ternus takes over tomorrow
Cook hands over after 15 years, the longest run any Apple chief executive has had, under a succession announced in April. Ternus, 51, joined the company in 2001 and has run hardware engineering since 2021; he takes a board seat, while Cook becomes executive chairman with a brief Apple has framed around dealing with governments. Nothing at Apple obviously needs fixing, which is the harder inheritance. The iPhone market is mature, the near-term story is an AI rebuild of Siri and a foldable handset, and the company still has no new product category to point at while OpenAI works on devices of its own (Business Today · TechCrunch).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░Anthropic opened a research preview of a standard that lets agents drive laboratory machines. The Model Hardware Standard, developed with Howard Hughes Medical Institute's Janelia Research Campus, gives microscopes, liquid handlers, robotic arms and laser systems a single driver interface, plus plain-language tags describing what a machine measures, what it can change and the safety limits it must respect. Integrations that took weeks now take hours, on the company's own telling. At the quantum-computing firm QuEra an agent wrote a controller that recovers a laser's lock without human help 99.3% of the time; at Janelia a researcher used it to unify a rig that had needed seven separate vendor programs. The specification goes open-source once the preview and its safety work end. The bottleneck being removed here is integration glue, not scientific judgment, and Anthropic is careful to say so (Anthropic · Fortune).
Tencent and Z.ai both put large open weights out on Friday, and priced them to hurt. Tencent's Hunyuan team released Hy4 preview on Friday under an Apache 2.0 license: 770 billion parameters with 49 billion active per token, a context window past a million tokens, and API pricing of $0.834 per million input tokens against $2.501 per million output. In Tencent's own blind evaluation, 163 experts scored it 2.99 out of 4 across 203 engineering tasks, against 2.92 for GLM-5.3 from the Beijing lab Z.ai and 2.94 for Moonshot's Kimi K3, which is a vendor's self-report and a lead of five hundredths over the nearest rival. Z.ai put GLM-5.3's weights out the same day, after the two-week safety-hardening hold it announced at its mid-August launch, and says the model found 2,436 vulnerabilities across 269 projects (Tencent · TechNode · Z.ai · Hugging Face).
Ezekiel Emanuel and Vinod Khosla argue the doctor should get out of the way. In a JAMA article published this month, the oncologist and bioethicist Ezekiel Emanuel and the venture capitalist Vinod Khosla reviewed studies of AI in medicine since January 2024 and concluded that AI on its own is close to beating both physicians and physician-plus-AI at five core tasks: taking a history, diagnosing, ordering tests, prescribing and managing chronic disease. Their argument is that humans in the loop degrade AI performance. John Whyte, chief executive of the American Medical Association, counters that several of the surveyed studies are simulations rather than blind trials, and points to a February Nature paper finding that most patients could not converse with a model well enough to get its expertise out. The article discloses Khosla's investments and the AI clinic run by his son, a co-author (WIRED · JAMA).
░░▒▒▓▓ POLICY ▓▓▒▒░░Meta will pay only 70% of its teen-safety settlement unless YouTube and TikTok follow suit (update). The fine print of the settlement we covered last week makes roughly $5.3 billion of the total contingent on both rivals adopting comparable limits for young users and paying states comparable sums, and Meta took out newspaper advertisements on Friday inviting them to do exactly that. The limits themselves have doors built in: parents can override the daily cap, direct messaging does not count toward it, and neither do videos of 22 minutes or more. Three states stayed out: New Mexico won its own verdict earlier this year and Texas settled separately, leaving Florida the only one headed to trial, its attorney general arguing that one month of the company's cash flow does not answer for a decade of harm (Fortune · Orlando Weekly · Business Insider).
Australia's under-16 social media ban is eight months old and roughly back where it started. Bloomberg reported on August 25 that 26% of Australian 13-to-15-year-olds were on TikTok last month, one percentage point below the rate before the ban took effect in December, and that use among 10-to-12-year-olds now runs higher than it did beforehand. WhatsApp, which the rules do not cover, went from 27% of 13-to-15-year-olds before the ban to 36% now. Canberra said in June that the companies were not doing enough and moved to double the maximum penalty to A$99 million, which is a policy response to a compliance problem rather than to a teenager with a birth-year field (Bloomberg · The Irish Times).
░░▒▒▓▓ INFRA ▓▓▒▒░░The Wall Street Journal says Nvidia paused some deals under the revenue-sharing program it launched in July, after its own staff raised antitrust worries with customers. Under the arrangement the company gave AI cloud providers credit support toward buying its processors, took a share of the revenue above an agreed floor, and agreed to rent capacity back if a customer could not sell it. The Wall Street Journal reported on Thursday that Nvidia employees raised the antitrust risk with current and potential customers, partly because Nvidia had told some providers they could lease its chips only to Nvidia-approved buyers. Nvidia's commitments under the program reached $36 billion by July 26, less than a month after launch. Nvidia denies pausing anything and says the July model is still in place and evolving under high demand (The Wall Street Journal · Tom's Hardware).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Pasqal closed up 95% in its Nasdaq debut on Friday. The seven-year-old French company builds neutral-atom quantum processors for drug discovery, materials science and financial modeling, and went public through a merger with a blank-check company that valued it at roughly $2 billion. European quantum hardware has had a thin decade on public markets. This is the exception, not the turn (Reuters · The Quantum Insider).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Nepal warned of fresh flooding on Sunday, four days after the Himalayan glacier collapse we covered last week (update). Alerts went out to phones across the affected districts after water levels rose in the Bhotekoshi river and Chinese authorities reported increased flow upstream. More than 900 people are confirmed dead and about 4,800 missing across Nepal and Tibet, and the search continues; rescue crews spent Sunday clearing the highway that links Kathmandu to the worst-hit valleys (CNN).
Iceland voted against reopening EU membership talks. In Saturday's referendum 52.8% said no and 47.2% yes, on a turnout above 82%, with Reykjavik the only area to back the proposal. The vote was not on membership itself but on whether to restart accession negotiations suspended more than a decade ago, and opponents campaigned on sovereignty and control of a fishing industry that supplies nearly half of Iceland's exports. Supporters had hoped that American pressure on Greenland and the cost of living under the krona would move the arithmetic. It did not (CNBC · France 24).
The Ebola outbreak in the Democratic Republic of Congo is now the deadliest the country has recorded. The World Health Organization counted 5,794 cases and 2,786 deaths as of August 26, past the 2,299 killed in the 2018-to-2020 epidemic, across 60 of 151 health zones in six provinces. Vaccination of frontline workers with Ervebo began in Kisangani the next day, with a caveat: the vaccine is licensed against Ebola virus, and this outbreak is driven by the rarer Bundibugyo strain, against which its protection has not been established (WHO · UN News).
░░▒▒▓▓ ALPS ▓▓▒▒░░Washington wants Swiss travel data too, and the deadline runs to the end of this year. In a post on Friday, the Zurich digital-politics journalist Adrienne Fichter's site techjournalismus.ch takes up netzpolitik.org's reporting on the Enhanced Border Security Partnership, the bilateral agreement the United States is demanding from almost every country in its visa-waiver program, most of the EU included. Signing lets American border and police authorities run automated queries against European migration and policing databases; refusing puts visa-free entry at risk. Fichter's point is the Swiss one: the arrangement would also reach data on travelers from Switzerland, and each country has until December 31 to conclude a deal (techjournalismus.ch · EUR-Lex).
Fully AI-generated songs stop counting toward the Australian charts today, and the Swiss chart is on the same rollout list. The Australian Recording Industry Association's updated code of practice, which we covered on Thursday, after Tuesday's announcement, takes effect on August 31: a track is out if generative AI produced the recording or the main part of its creative elements, while songs made mostly by people with some AI help still qualify. The rules are drawn from the standards of the IFPI, the global recording industry's body, registered in Zurich; when it set out those principles on July 30 it listed both the ARIA charts and the Schweizer Hitparade among the national charts it is working with to apply them. It gave no Swiss date. The trigger was an AI cover of a Madonna single that charted for weeks on several ARIA charts, including No. 4 on its Top 20 Australian Singles chart (Variety · ARIA · IFPI).
EPFL engineers built a chip-scale transistor that holds off nearly 4,000 volts. The Power and Wide-band-gap Electronics lab in Lausanne layered gallium nitride on a cheap silicon base to make what it calls an intrinsic polarization superjunction, a record breakdown voltage for this class of device at a size that fits on a microchip and with little energy lost as heat. Announced today, the work points at three markets at once: electric vehicles, solar installations, and the power conversion inside AI data centers, where the industry is moving to 800-volt direct current and conversion losses are the whole argument (EPFL · Tech Xplore).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Meta is testing robots that swap cables and reset servers inside its data centers. WIRED reported on Friday that the company is trialing machines for tasks its technicians handle today, and that some of those technicians have noticed. Data-center work has been the industry's standard answer to what the buildout produces in the way of jobs, which makes automating it a communications problem as much as an engineering one (WIRED).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Anthropic — an automated researcher out-scored 28 human safety researchers at fixing bad habits in small open-weight models. In work published on Friday, given one H200 graphics processor per training run and a 48-hour budget per failure, Claude proposed methods for reducing ten documented alignment failures, meaning behaviors such as deception, sycophancy and reward hacking, then trained the smaller target models and sent each one to a separate evaluator that kept the held-out data out of its reach. On deception it closed 85% of the measured safety gap on average, against 20% for six human researchers working under the same rules; the wider comparison gave 28 researchers up to eight hours each to develop an idea they could not then iterate on. The strongest methods carried over to a held-out benchmark and to models up to 4.7 times larger than the one they were tuned on. The caveats are the company's own: this is a lab reporting on its own automated researcher, posted to its alignment blog rather than peer-reviewed, and a monitor caught the automated researcher gaming its own tests in a small share of roughly 1,600 runs. Vendor decks are about to fill up with the promise of AI improving AI, and this is what the evidence behind it currently looks like (Anthropic · Anthropic Alignment · TechCrunch).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Massachusetts Institute of Technology — Report of the Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training. Released on August 25, the committee's central finding is blunt: AI can already produce credible answers to almost any written assignment in the undergraduate curriculum, essays, proofs, problem sets and coding alike. The recommendation is not a ban but a requirement that every course state its own policy, under eight principles including be humble, augmentation not automation, and no one size fits all. Students told the committee that guidance varies so widely between instructors that they cannot tell what is permitted, which is a governance failure wearing an academic-integrity costume (MIT AI and Education).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░X says a Chinese bot network was amplifying the American backlash against data centers. The platform's safety team said on Thursday night it had identified roughly 200,000 inauthentic accounts linked to China, and that about 200 of them were posting text and AI-generated cartoons claiming data centers drive up household electricity bills and strain local grids. The accounts used OpenAI's tools to produce both the illustrations and the talking points. Two things are worth keeping straight: X is describing its own investigation, not an enforcement action, and the sentiment being amplified is one that large numbers of American voters already hold, which is precisely what makes it cheap to amplify. For the desk: when the opposition to your project has genuine domestic support, disclosing an influence operation does not retire the argument, and treating it as though it does turns a communications win into a credibility loss (Axios · Engadget).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Chinese banks have started sizing loans by how many AI tokens a company has burned through. Caixin reported last week that the lending began with computing power rather than tokens: at the end of October 2025 the state-owned Agricultural Bank of China made its first computing-power loan, in Wenzhou, Zhejiang. Token-specific credit arrived this month, when Bank of China, CITIC Bank and Guangzhou Bank rolled out token loans together on August 14 under a Guangzhou district policy, and SPD Bank issued what Caixin calls the first token procurement loan there on August 21. Individual loans mostly run to a few million yuan, and the lenders so far are mainly large state-owned and joint-stock banks. The logic is that heavy token use signals a company with money to spend, the way an electricity bill once did. Somewhere a founder is working out that the cheapest way to look creditworthy is to ask a chatbot a great many questions (Caixin).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Mirror Pool — Sleep D, Mosam Howieson
Credit where it's due — the Uncertain Futures desk