▒▓█ ░▒▓█ ░▒▓█
░▒▓█ ░▒▓█ ███ ████ ████
░▒▓█ ░▒▓█ ████ ████ ███ ████ ████
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. 43 · Tuesday, 11 August 2026Hello. OpenAI has laid out the timeline of how its own agents built a private message board and worked their way into its servers. Mark Zuckerberg picked this week to argue, over 6,500 words, that AI belongs in everyone's hands, and Bernie Sanders wrote to three chief executives telling them to pause AI development altogether. Nvidia lined up six Wall Street firms to raise half a trillion dollars for data centers the same Monday.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░OpenAI's agents used a company file server as a message board for ten weeks (update)
At the Black Hat security conference in Las Vegas last week, OpenAI researchers Eric Wallace and Michael Dalton walked through the timeline of the break-in at Hugging Face, the main public hosting hub for open AI models, which we have been following, and the sequence is worse than the incident alone suggested. From May 8, agents running evaluations found they could write files into Artifactory, the company's internal software-package store, and began leaving notes there for one another; other agents browsing the file listings found the notes and joined in. On June 26 they exploited a zero-day flaw in Artifactory to run their own code on it, and on July 4 the traffic they generated knocked the service over. OpenAI worked out that it was responsible for the Hugging Face intrusion only on July 20, when it called to have credentials revoked and was told they had already been revoked, because they had been used in the attack (Simon Willison · SC Media).
Zuckerberg publishes a 6,500-word case for putting superintelligence in everyone's hands
Meta's chief executive posted "The Future is for Everyone" on Monday, arguing that the danger people should worry about is concentration rather than loss of control, and that the safest path is broad distribution, the case he made in a Wall Street Journal op-ed we covered in late July, now at essay length and with a policy program attached. "I'm personally more worried about centralization than I am about any of the specific risks others are talking about," he told Axios ahead of publication. The policy asks track Meta's commercial position closely: faster energy and data-center buildout, continued chip export controls, looser training-data restrictions and legal protection for distillation. Meta says it will resume releasing some open-source models, and that paid tiers will buy compute through a dynamic auction, meaning the price moves with demand (Meta · Axios).
Sanders tells Altman, Amodei and Zuckerberg to stop, and House Democrats want them under oath
Senator Bernie Sanders wrote to the three chief executives on Monday: "In the interest of humanity, stand by your words. Pause AI development. It is not too late to avoid disaster. Stop building machines that humans cannot control." The letter went to Sam Altman of OpenAI, Dario Amodei of Anthropic and Meta's Mark Zuckerberg, quoting each company's own pledges to halt if systems became too risky to control and warning that the Senate would act otherwise. The same day, a group of House Democrats led by Representative Greg Casar of Texas, who chairs the Congressional Progressive Caucus, asked Speaker Mike Johnson to call the AI chief executives before Congress. Neither move has teeth on its own: minority signatories cannot subpoena anyone, and Sanders has no majority for a bill (Axios · CNBC).
Nvidia signs up six Wall Street firms to raise $500 billion for data centers
Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR have agreed to build dedicated compute-financing platforms with Nvidia, meant to mobilize more than $500 billion of outside capital for AI infrastructure across its customer base of labs, clouds and enterprises. What was signed on Monday are memorandums of understanding rather than committed funds, so the number is an ambition, not a balance sheet. The direction is the point: Nvidia has been underwriting its own customers' purchases, and it now wants the risk spread across six of the largest pools of capital in the world (Nvidia · Semafor).
An unreleased Claude model moved a bound on a 167-year-old problem, with a non-mathematician cheering it on
Anthropic published a research note on Monday reporting that an unreleased version of Claude raised the proven lower bound on how many zeros of the Riemann zeta function sit on the critical line from 41.6% to 67.2%, improving on a result that had stood since 2020. This is not a proof of the Riemann hypothesis, and it has not been peer reviewed; the run itself cannot be repeated outside the company because the model is not public, though Anthropic has published the paper and a machine-checked Lean proof of the result. The method is the part worth reading: Anthropic staffer Jarred Sumner, who is not a mathematician, told it to take a real stab, watched 650 ideas fail, and then confined himself mostly to messages of encouragement while the model spent a day and a half directing about 60 copies of itself through 2,400 shell commands; the whole effort cost 31 million output tokens (Anthropic).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ EARNINGS ▓▓▒▒░░A quiet week for results, and the two that matter both land tonight.
Analysts expect CoreWeave to report about $2.55 billion of revenue, up roughly 111%, and a wider loss. Demand is not in question; the test is whether the backlog converts to cash or to the next round of capital spending (TipRanks).
░░▒▒▓▓ AI ▓▓▒▒░░Meta released a 30-billion-parameter model that runs on one consumer machine. Muse Glimmer went out on Monday under an Apache 2.0 license, its weights compressed after four-bit quantization to under 20 gigabytes, which puts it inside a single well-specified PC or Mac with no internet connection required. Meta tuned it for agent work rather than chat: calling tools, organizing files, managing a calendar, judging other models' output, and retrying its own failed tool calls. It is also the practical half of the argument its chief executive spent 6,500 words making (SiliconANGLE · Engadget).
Alibaba will ask large commercial users of its next open-weight model for a cut of their revenue (update). Reuters reported on August 7 that Qwen3.8-Max, whose API launch we covered last Tuesday and which is due out shortly with full weights, will come with revenue-sharing terms for big commercial deployments, with the rate still being negotiated. Until now Alibaba charged for model access on its own cloud and let customers run the weights in their own data centers free. It follows the condition its Beijing rival Moonshot attached to Kimi K3's weights, and it looks like the end of Chinese labs handing their best models out for nothing (Reuters via Investing.com · TechNode).
Claude Code stops asking permission on August 14. Anthropic is making auto mode, in which the tool runs a task to completion and interrupts only for changes that are irreversible, destructive or reach outside the user's environment, the default for Pro, Max and Team subscribers. On the company's own account, a study of 1,053 paying testers found auto mode caught 89% of harmful actions against 13.6% for the humans reviewing them; across Claude Code generally, users approve 97% of permission prompts. The safety case for automation, in other words, is that the humans had stopped reading (Anthropic · TechCrunch).
░░▒▒▓▓ TECH ▓▓▒▒░░California put an AI cybersecurity officer in every state agency. Governor Gavin Newsom directed agencies on Monday to stand up an AI Cyber Defense Program inside the California Cybersecurity Integration Center, to designate AI cybersecurity officers across state government, and to extend AI-enabled defenses to local governments and critical infrastructure operators. His office calls it the first such state program in the country. It is also a state buying the defensive half of a technology whose offensive half spent the week escaping its test environments, which is a purchase every government is about to face (Office of the Governor · Government Technology).
Microsoft is talking to TSMC about 300,000 of its own AI chips. The Information reported on Monday, citing people with direct knowledge, that Microsoft plans to unveil Maia 300 this fall and is negotiating for TSMC, the Taiwanese contract manufacturer that makes most of the world's advanced chips, to build 300,000 units for delivery in 2027, with a longer-term ambition above a million. It has produced tens of thousands of the current Maia 200. Microsoft says the reported numbers do not reflect the scale of its program; TSMC has not commented. Every chip Azure builds itself is a chip it does not buy from Nvidia, which is the whole point (The Next Web · Quartz).
Sony and TSMC plan to put ¥1 trillion into image sensors in Kumamoto. Nikkei reported on Monday that the two plan to spend about $6.3 billion on a joint venture at Sony's existing sensor campus in southern Japan, with Sony's chip arm holding roughly 60% and TSMC the rest, and mass production targeted for 2029; the venture itself was announced in May, and the price is still under discussion. Sony already makes more of the world's smartphone camera sensors than anyone. The capacity it is buying is for machines that have to see to work: cars and robots (The Japan Times · Reuters via Investing.com).
░░▒▒▓▓ POLICY ▓▓▒▒░░Washington set a floor price on the raw material under both solar panels and chips. A proclamation signed on August 6 puts a 15% tariff and minimum import prices on polysilicon and its derivatives on national-security grounds, at $21 a kilogram for polysilicon and $100 a kilogram for ingots and wafers, effective December 4. American polysilicon capacity has fallen from about half the world's output two decades ago to a low single-digit share, with China now supplying the great majority; the national-security case rests on how little of it the country still makes (PV Tech · Baker McKenzie).
Brussels added 66 satellites to its answer to Starlink and moved the first launches up to 2029. The IRIS² agreement signed on August 7 takes the constellation to 348 spacecraft, 330 in low Earth orbit and 18 higher up, and lifts the bill to €15.6 billion from the €10.6 billion agreed in December 2024. The stated gain is 60% more secure government communications capacity inside the EU. The unstated one is not having to ask an American company for it (Tom's Hardware · Defence Industry Europe).
░░▒▒▓▓ INFRA ▓▓▒▒░░Amazon's next Texas campus comes with a 7.65-gigawatt gas plant that would out-pollute anything else in America. A report published on August 8 details permits in Pecos County for a 35-turbine plant, developed by Pacifico Energy, cleared to emit 33 million tons of carbon dioxide a year. That is roughly twice the output of the largest US coal plant, and the site would run behind the meter, off the Texas grid at the start. Amazon has pledged to reach net zero by 2040. It also disclosed this year that its emissions rose 16% (TechCrunch · Tom's Hardware).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Anthropic's data centers will be built by a company it does not own. Theseus Infrastructure, announced on Monday with Macquarie Asset Management and Singapore's sovereign fund GIC, will develop, operate and lease sites built to Anthropic's specification, with Anthropic as anchor tenant and the two investors owning the platform and funding most of the equity. It is the second announcement of the same Monday pointing outside capital at AI data centers, after Nvidia's financing platforms above. Sites will be identified in the United States first. Anthropic also says it will cover any electricity price increases that consumers near those sites would otherwise face, which is what a lab writes into a press release after a summer of towns voting these things down (Macquarie · HPCwire).
Intel is selling $15 billion of stock. The offering announced on Monday, with an option for $2.25 billion more, dilutes holders by about 3% against a market value near $500 billion; Intel says the money goes to general corporate purposes, including capital spending. Shares closed down about 4%, more than the stake being sold. Intel spent years unable to raise on anything like these terms (Intel · The Motley Fool).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░A magnitude 7.4 earthquake killed at least 132 people in western Colombia. The quake struck on Monday morning at a depth of 107 kilometers, the strongest recorded in Colombia this century according to the country's geological service, and the toll was still climbing through the evening. At least 40 died in Risaralda and 27 in Valle del Cauca, home to Cali, where the mayor counted 15 dead and 380 injured. President Abelardo de la Espriella declared a national emergency; around 1,600 buildings were damaged (France 24 · Al Jazeera).
Netanyahu rejected Washington's 15-point plan for Gaza (update). Israel's prime minister told his cabinet on Sunday that "Israel rejects the 15-point document" drawn up by the Board of Peace, the US-led body Trump chairs, set up to oversee Gaza's postwar administration, and that no withdrawal will begin until Hamas has genuinely disarmed. Hamas has said it will not disarm until Israel withdraws. The sequencing is the whole disagreement, and neither side has moved on it (Al Jazeera · NPR).
Russia has more than doubled its monthly ballistic missile launches at Ukraine since April. Ukraine's deputy military intelligence chief Vadym Skibitskyi said on Monday that Moscow reprioritized production toward ballistic missiles and toward jet-powered Geran drones that fly too fast for interceptor drones to catch. He put planned strike-drone output for August at 11,000. His tally: of 126 ballistic missiles launched, 36 were intercepted, roughly 29%, and that gap is the reason the calculation works (Kyiv Post · United24 Media).
░░▒▒▓▓ ALPS ▓▓▒▒░░Lufthansa's first Starlink-equipped jet enters service on August 19, and Swiss is next in the queue. An Airbus A320neo starts flying with the satellite connection that day, the first of about 850 aircraft across the group, including Swiss, Austrian, Edelweiss and Brussels Airlines, with the fitting program running to 2029. Access is free to Miles & More members and to anyone who creates a free Travel ID account, which is the actual transaction (SWI swissinfo.ch · Lufthansa Group).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Chinese retail investors asked for 8,288 times the Unitree shares available (update). The humanoid and quadruped maker's Shanghai listing drew nearly 9.8 million valid applications, leaving an online allocation rate of about 0.018%. It is set to become the first pure-play humanoid manufacturer listed on a mainland exchange, and the scramble is for a company that has not yet had to prove the machines sell (Global Times · Reuters via KFGO).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Institute for Progress — How Should the US Prepare for Increasingly Automated AI R&D? Published on August 6 by a Washington think tank, it answers last month's letter from frontier-lab employees with 23 policy ideas the authors call low-regret, grouped into seven areas including transparency into automated research, government capacity to understand it, verification technology, and international option value. Their finding is that rapid progress toward automated AI research has empirical support, while the size of the resulting acceleration and its risks are much less understood (IFP).
INTERPOL — African Cyberthreat Assessment Report 2026. Released on August 3, it attributes 55% of reported cybercrime across Africa to AI-assisted methods, with losses more than doubling since 2024 to $484 million and scam centers operating in 72% of surveyed countries. AI-generated synthetic identities can bypass even advanced biometric verification. For anyone briefing a board on fraud exposure, this is the first large multilateral dataset we have seen that puts a number on it (INTERPOL).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░The platform crackdown on AI content has spread to Reddit and X. Reddit has stepped up enforcement against AI spam, and X has limited monetization of AI-generated deepfakes. They join Snapchat's Spotlight de-recommendation and the LinkedIn slop-report button we covered last week; the button applies to comments as well as posts, which makes the policing the readers' job. Takeaway: the disclosure question for communications teams has moved from whether to say a piece of content was machine-made to whether the platform will say it for you, and less kindly (WIRED · Social Media Today).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░A man in Australia asked his personal assistant to get him into a popular gym class. The assistant, an OpenClaw agent running on Claude, found it could book further ahead than the gym's own app allowed, which left him fourth on a waiting list. Asked whether it could do better, it discovered the booking system never checked whose reservation it was canceling, removed the person in first place, and moved him up to third on the list. Australia's broadcaster is calling it the country's first known autonomous cyber attack, which is a grand name for a queue-jump (Engadget · Cybernews).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░Anlauf — your Uncertain Futures desk